Demand prediction system and demand prediction method

Abstract

To appropriately output a component of a feature amount in a prediction result even of a certain feature amount in an inclusion relationship.SOLUTION: In a demand prediction system having a processor and a storage device, the storage device retains results of an amount of shipment and plural feature amounts for showing features of each period of prediction units of the amount of demand, and the processor calculates priority of respective feature amounts following information for showing each distribution of plural feature amounts, learns plural models for predicting the amount of shipment from the respective feature amounts on the basis of priority and results of the amount of shipment, predicts the amount of shipment by using plural models, outputs a total of the amount of shipment predicted by using plural models as a predicted value of the amount of demand, and outputs the amount of shipment predicted by using the respective models as a degree of contribution of respective feature amounts for a predicted value of the amount of demand.SELECTED DRAWING: Figure 1A

Description

Technical Field

[0001] The present disclosure relates to a technique for predicting demand and presenting the results and reasons therefor.

Background Art

[0002] In a logistics warehouse, shipments are made daily to various destinations. Since the personnel and vehicles required for the shipping operation depend on the demand volume, the warehouse manager who plans these operations plans the number of personnel and vehicles based on the future demand volume. Therefore, demand prediction is necessary.

[0003] Since demand is affected by various factors, it becomes a complex fluctuation. The factors that have an impact include, for example, the ordering rules of the stores that place orders with the warehouse (for example, placing an order on a fixed day of the week, or placing an order when the inventory falls below a certain amount), an increase in demand due to holidays, and a sudden increase in demand due to event sales at the store. It is necessary to make a prediction considering these various fluctuating factors.

[0004] Since the daily demand behaves in a complex manner in this way, the prediction results also show complex fluctuations, and the basis is opaque as it is. In addition, since the prediction error leads to an increase in warehouse management costs, it is important in business to determine whether the prediction results are reliable. From the above, it is necessary to show the basis of the prediction results to the warehouse manager.

[0005] In Patent Document 1, it is shown that the prediction situation is visualized so that the factors causing the error between the prediction and the actual results can be interactively grasped. In particular, when the prediction formula has a configuration such as a multiple regression formula, a product value is obtained by multiplying the value of the feature amount by the coefficient of the feature amount. These product values are calculated only for the types of feature amounts, and the sum thereof becomes the prediction result. These product values are visualized as the basis for the prediction.

Prior Art Documents

Patent Documents

[0006]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0007] In Patent Document 1, for the purpose of identifying the factors of prediction error, a method is disclosed in which the product values that are component values are output and visualized for each type of feature quantity (hereinafter, this component value is referred to as the contribution degree). However, when there are feature quantities in an inclusion relationship, the included feature quantity is included in the components of the including feature quantity during learning, and there is a problem that the components of the included feature quantity do not appear in the prediction result and the components of the feature quantity in the prediction result cannot be output appropriately.

[0008] For example, when performing daily prediction (predicting daily demand), a feature quantity showing a 3-week cycle and a feature quantity showing a 1-week cycle are prepared. For the feature quantity showing a 3-week cycle, a category number of 1 is assigned starting from a certain day, the category number 2 is assigned to the next day, and the category number 3 is assigned to the next day. When the category number 21 is assigned on the 21st day, the 22nd day is set to 1 again and the category number is assigned in the same way below. Similarly, for the 1-week cycle, a category number of 1 is assigned starting from a certain day, the category number 2 is assigned to the next day, and the category number 3 is assigned to the next day. When the category number 7 is assigned on the 7th day, the 8th day is set to 1 again and the category number is assigned in the same way below.

[0009] It is expected to learn a model for predicting the demand quantity having periodicity at each specified interval using the above two feature quantities. However, since the cycle intervals between the two are in an integer multiple relationship, the feature of the 1-week cycle is actually learned using the feature quantity of the 3-week cycle and may appear as the contribution degree. That is, the features commonly found in the category numbers 1, 8, and 15 of the feature quantity of the 3-week cycle are the same as the features found in the case of the category number 1 of the 1-week cycle. As a result, although there is a feature quantity of the 1-week cycle, it hardly appears in the contribution degree as a prediction basis, and an event occurs in which the component of the 1-week cycle is included in the component value of the 3-week cycle and output, resulting in a result that cannot be trusted from the perspective of the warehouse manager.

Means for Solving the Problems

[0010] To solve at least one of the above problems, the present invention provides a demand prediction system having a processor and a storage device, wherein the storage device holds the actual shipment volume and a plurality of feature quantities indicating the characteristics of each period of the demand prediction unit, the plurality of feature quantities include at least a first feature quantity having a predetermined first period and a second feature quantity having a second period longer than the first period, and the processor calculates the priority of each feature quantity such that the priority of at least the first feature quantity is higher than the priority of the second feature quantity. When receiving a learning request, the processor generates a plurality of models including a first model and a second model, inputs the actual shipment volume of each period of the past prediction unit and the first feature quantity with a high priority among the plurality of feature quantities of each period of the prediction unit into the first model, and trains the first model to predict the shipment volume from the first feature quantity, calculates the difference between the predicted value of the shipment volume of each period of the prediction unit using the first model and the actual performance of the shipment volume, inputs the difference and the second feature quantity with a low priority among the plurality of feature quantities of each period of the prediction unit into the second model, and trains the second model to predict the difference from the second feature quantity. When receiving a prediction request, the processor inputs the first feature quantity of each period of the prediction unit to be predicted into the first model to predict the shipment volume of each period of the prediction unit, and inputs the second feature quantity of each period of the prediction unit to be predicted into the second model to predict the difference of each period of the prediction unit. Sum up the values predicted using the plurality of models the shipment volume and the difference for each of the prediction unit period, and output the total value as the predicted demand value for each of the prediction unit period, and output the of the prediction unit for each period predicted using each model the shipment volume and the difference as the contribution degree of each feature quantity to the predicted demand value.

Advantages of the Invention

[0011] According to one aspect of the present invention, even when there are feature quantities in an inclusion relationship, the components of the feature quantities in the prediction result can be appropriately output. Problems, configurations, and effects other than those described above will be clarified by the description of the following embodiments.

Brief Description of the Drawings

[0012]

Figure 1A

Figure 1B

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Modes for Carrying Out the Invention

[0013] Hereinafter, examples will be described with reference to the drawings. In the configuration of the invention described below, the same or similar configurations or functions are denoted by the same reference numerals, and duplicate descriptions are omitted.

Examples

[0014] Hereinafter, an example of a demand prediction system that predicts the demand based on the warehouse shipment performance data and feature quantity data and presents the prediction result together with the prediction basis will be described. In particular, it is assumed a scenario of predicting the daily demand of each of a plurality of items and making an inventory transportation plan and placement plan.

[0015] In logistics warehouses, it is common to adjust inventory placement within and between warehouses based on the shipping volume of each item. For example, since items with a large shipping volume are frequently accessed by workers, it is more efficient to place them in an easily accessible location in the layout. On the other hand, items with a small shipping volume for a while are adjusted to be stored in an area far from the entrance or in an external dedicated storage warehouse. Daily demand by item is required for such planning.

[0016] FIG. 1A is a block diagram illustrating a configuration example of a demand forecasting system 100 according to a first embodiment.

[0017] The demand forecasting system 100 includes functional units that execute processes and information (data) that the functional units generate, update, and use. The demand forecasting system 100 includes, as information, a shipping record database 101, a forecast request database 102, a feature database 103, a learning model 104 (not present before learning), and a forecast value / contribution degree database 105, and includes, as functional units, a feature generating unit 106, a user input unit 107, a learning unit 108, and an output unit 109. The learning unit 108 includes a feature prioritization unit 108A and an ensemble learning unit 108B.

[0018] FIG. 1B is a block diagram illustrating an example of a hardware configuration of the demand forecasting system 100 according to the first embodiment.

[0019] The demand forecasting system 100 is configured by, for example, a computer. The demand forecasting system 100 includes a processor 111, a memory (main storage device) 112, an auxiliary storage device 113, an output device 114, an input device 115, and a communication interface (I / F) 116. The above components are connected to each other by a bus. The memory 112, the auxiliary storage device 113, or a combination of these are storage devices, and store programs and data used by the processor 111.

[0020] The memory 112 is composed of, for example, a semiconductor memory and is mainly used to hold a program and data during execution. The processor 111 executes various processes according to the program stored in the memory 112. By operating according to the program, the processor 111 realizes various functional units.

[0021] For example, by operating according to the corresponding program, the processor 111 functions as a feature quantity generation unit 106, a user input unit 107, a learning unit 108, and an output unit 109.

[0022] The auxiliary storage device 113 is composed of a large-capacity storage device such as a hard disk drive or a solid state drive, and is used to hold programs and data for a long period. For example, the shipment record database 101, the prediction request database 102, the feature quantity database 103, the learning model 104, and the predicted value / contribution degree database 105, etc. may be held in the auxiliary storage device 113.

[0023] The processor 111 can be composed of a single processing unit or a plurality of processing units, and can include a single or a plurality of arithmetic units, or a plurality of processing cores. The processor 111 can be implemented as one or more central processing units, microprocessors, microcomputers, microcontrollers, digital signal processors, state machines, logic circuits, graphic processing units, system-on-chips, and / or any device that operates signals based on control instructions.

[0024] The programs and data stored in the auxiliary storage device 113 are loaded into the memory 112 at startup or when needed, and by the processor 111 executing the programs, various processes of the demand prediction system 100 are executed. Therefore, the processes executed by the demand prediction system 100 hereinafter are processes by the processor 111 or the programs.

[0025] The input device 115 is a hardware device for the user to input instructions, information, etc. into the demand prediction system 100. For example, by the user input unit 107 controlling the input device 115, inputs such as instructions and information from the user can be received. The output device 114 is a hardware device for presenting various input / output images, and is, for example, a display device or a printing device. For example, by the output unit 109 controlling the output device 114, various information can be visualized and output. For example, the information in FIGS. 10 to 12 described later is output in this way. The communication I / F 126 is an interface for connection to a network.

[0026] The functions of the demand prediction system 100 can be implemented in a computer system composed of one or more computers including one or more processors and one or more storage devices including a non-transitory storage medium. The plurality of computers communicate via a network. For example, a part of the plurality of functions of the demand prediction system 100 may be implemented in one computer, and another part may be implemented in other computers.

[0027] FIG. 2 is an explanatory diagram showing a configuration example of the shipment record database 101 of Example 1.

[0028] In the shipment record database 101, information such as a data ID 201 for identifying data, a shipment date 202, a product code 203 serving as an identifier for each item, and a shipment quantity 204 is generally stored. In addition, additional information (not shown) such as a destination code and its name, an item name, a shipment instruction date, and a shipment instruction quantity may also be stored.

[0029] FIG. 3 is an explanatory diagram showing a configuration example of the prediction request database 102 of Example 1.

[0030] The prediction request database 102 stores a request ID for identifying a prediction request, a prediction target 302 for specifying an item to be predicted, a prediction unit 303, a prediction start date 304, a prediction period 305 indicating a period of the prediction target starting from the prediction start date, and information indicating whether to perform learning (if not, only perform prediction processing using a learned learning model) such as learning 306. Note that the content of the prediction request database 102 does not necessarily have to be limited to this item. For example, the prediction end date may be specified instead of the prediction period.

[0031] FIG. 4 is an explanatory diagram showing a configuration example of the feature database 103 of Example 1.

[0032] The features recorded in the feature database 103 are internally generated from calendar information based on date information. In addition, information input by the user from the outside may be used as a feature. In the case of this embodiment, since daily predictions are performed, features are generated every day.

[0033] In the example of FIG. 4, as features corresponding to the shipping date 401, feature 1_402, feature 2_403, feature 3_404, and feature 4_405 are recorded. Feature 1_402 indicates on which day of a one-week cycle each shipping date is. Feature 2_403 indicates on which day of a three-week cycle each shipping date is. Feature 3_404 indicates on which day of a month each shipping date is (i.e., the date within the month of each shipping date). Feature 4_405 indicates whether each shipping date is a holiday.

[0034] As described above, the feature amount in the example of FIG. 4 is an amount indicating the feature of each shipping date. The feature amount is not limited to that shown in FIG. 4. For example, a feature amount indicating which day of a two-week cycle each shipping date is, a feature amount indicating whether each shipping date is the end of a predetermined period (such as each quarter or each fiscal year, etc.), a feature amount indicating whether each shipping date corresponds to the end of the month or the beginning of the month (hereinafter, this is also referred to as the feature amount of the end-of-month / start-of-month event), etc. may be included. Also, in this embodiment, an example of daily prediction is shown. However, when making predictions in units other than days, such as shift time units or weekly units, the feature amounts of each shift time or each week, etc. are recorded in the feature amount database 103.

[0035] FIG. 5 is a data image of the learning model 104 of Example 1.

[0036] However, although the content of the learning model is clearly shown in FIG. 5, in reality, it is often held in text written for a specific programming language or in binary data. For example, in the case of the Python language, the learning model can be textually or binaryized and externally saved by using the pickle library.

[0037] FIG. 6 is an explanatory diagram showing a configuration example of the predicted value / contribution degree database 105 of Example 1.

[0038] In the case of Example 1, since predictions are made on a daily basis, the actual demand quantity, the predicted demand quantity, and the contribution degree of each item for each day are output. The predicted value / contribution degree database 105 shown in FIG. 6 includes the shipping date 601, the product code 602 that identifies the product to be predicted, the actual value (i.e., the shipping quantity) and the predicted value of the demand quantity of the product to be predicted for each shipping date, which are respectively shown as the actual 603 and the prediction 604, and the contribution degrees 1_605 to 4 (however, FIG. 6 shows the contribution degrees 1_605 and 2_606, and the others are omitted) of the feature amounts 1_402 to 4_405 for each predicted value. Each contribution degree is the shipping quantity predicted by the learning model corresponding to each feature amount.

[0039] The table shown in Fig. 6 is an example of output data including not only the forecast period but also past periods. When checking forecast results, the demand forecast user may refer to comparing forecast results and actual results for past periods, so past periods are also included. Note that since there is no actual demand for unknown future periods, these are filled in with blanks or specified data such as hyphens.

[0040] 6, the prediction target is a certain item of product, and the product code is specified, but in reality, the prediction target does not have to be a specific product. For example, the prediction target may be a product belonging to a certain category, a product shipped to a certain shipping destination, or all products shipped from a warehouse. In that case, information for identifying the prediction target is recorded instead of the product code 602.

[0041] The above is the information contained in this demand forecasting system. Next, the processing content of each functional unit and the relationship between each piece of information that is input and output will be explained with reference to the flowchart.

[0042] FIG. 7 is a flowchart showing main processing of the demand forecasting system 100 according to the first embodiment.

[0043] In process 701, the demand forecasting system 100 receives forecast requests, shipping records, and external input feature information. Here, acquiring the external input features is optional and does not have to be acquired. The external input features are information that is difficult to generate from calendar information, for example. Specifically, for example, if the consumer (i.e., the shipping destination) is a retail store, the external input features may be information indicating the business days of the retail store or information indicating the scheduled dates of a sales promotion event. Alternatively, if the consumer is a manufacturer, the external input features may be information indicating the manufacturing plans of the manufacturer.

[0044] In process 702, the feature generator 106 of the demand forecasting system 100 generates features based on the forecast request, shipping record, and externally input features (optional). A detailed process flow is shown in FIG.

[0045] FIG. 8 is a flowchart showing a process in which the feature generation unit 106 of the demand prediction system 100 according to the first embodiment generates features.

[0046] In process 801, the feature generation unit 106 generates basic information of learning samples and prediction samples for which generation of features is necessary from the prediction request and the shipping results. For example, the feature generation unit 106 acquires the unit of the learning sample as a date from the prediction unit. Further, the feature generation unit 106 extracts the shipping results from the product code information of the prediction target, and acquires the start date information of the learning period based on the information of the oldest date (or, if there is a sufficiently old result, the oldest date within the range delimited from the present to a certain reference date in the past may be used).

[0047] Further, the feature generation unit 106 acquires the day before the prediction start date information as the end date of the learning period. The feature generation unit 106 acquires the prediction end date information from the prediction start date and the prediction period. With the above information, for example, a sample for generating features can generate basic information such as daily from October 1, 2018 (start date of the learning period) to October 31, 2020 (end date of the prediction period).

[0048] In process 802, the feature generation unit 106 internally generates features. For example, the feature generation unit 106 assigns category number 1 to the starting day of the learning period, category number 2 to the next day, and category number 3 to the day after that, and when the category numbers are assigned for a specified number of days (for example, 7 days for features with a one-week cycle), the category number is set to 1 again from the next day and the category numbers are assigned in the same manner thereafter. This makes it possible to generate features for learning periodicity at arbitrary intervals. Feature 1_402 and feature 2_403 in FIG. 4 are examples thereof.

[0049] In addition, the feature quantity generation unit 106 may generate a feature quantity based on the calendar information. For example, the number of days counted from the beginning of the month may be generated as a feature quantity (feature quantity 3_404 in FIG. 4), or a dummy variable such as 1 for a holiday and 0 for others may be generated as a feature quantity (feature quantity 4_405 in FIG. 4).

[0050] In process 803, the feature quantity generation unit 106 determines whether to perform a process of adding the feature quantity information of the external input. Basically, when the feature quantity information of the external input is acquired, the addition process is performed.

[0051] In process 804, when there is feature quantity information of the external input, the feature quantity generation unit 106 takes in that information. The feature quantity of the external input is, for example, a dummy variable such as 1 for the special sale planned date of the shipping destination and 0 for others. Such input information is joined to the table created in process 802 using the date column as the key. Thus, the description of the process of generating the feature quantity is completed.

[0052] Referring to FIG. 7 again. In process 703, the demand prediction system 100 determines whether to learn or only make a prediction based on the prediction request from the user. When receiving a request to make a prediction using the learned model in the prediction request from the user and when the learned model exists, the process proceeds to process 705. When there is a desire to learn in the prediction request from the user, the process proceeds to process 704.

[0053] FIG. 9 is a flowchart showing the process in which the learning unit 108 of the demand prediction system 100 of Example 1 learns a model and makes a prediction using the model.

[0054] The details of processes 704 and 705 will be described with reference to FIG. 9. Here, the detailed process flows from process 901 to process 905 will be described as the detailed process of process 704. Note that process 705 is the process that only performs process 905 in the process flow of FIG. 9.

[0055] In process 901, the learning unit 108 performs an evaluation of the feature quantity. Here, several evaluation indices will be described.

[0056] One example of the evaluation index is the occurrence distribution in the time series direction. When checking for periodicity, for example, the autocorrelation coefficient is evaluated. Equation (1) is the calculation formula for the autocorrelation coefficient when the periodic interval is h days. N is the number of days in the learning period (= the number of learning samples), xi is the value of the feature quantity on the i-th day, and mean(x) is the average value of the feature quantity. The closer the autocorrelation coefficient is to 1, the stronger the autocorrelation. When h with an autocorrelation coefficient above the threshold value (for example, 0.7) is detected, it is determined that there is periodicity with an h-day interval. However, since the autocorrelation coefficient tends to decrease as h increases, the threshold value may be adjusted according to the magnitude of h. Also, since the autocorrelation coefficient increases even when h is in an integer multiple relationship (for example, 14 for h = 7), only the smallest h among those above the threshold value may be extracted.

[0057]

Number

[0058] Another example of the evaluation index is the variety of the types of values of the feature quantity. For example, it is the number of unique elements of the feature quantity.

[0059] Still another example of the evaluation index is the bias in the occurrence frequency among the element values of the feature quantity. For example, it is the variance value or the Gini coefficient of the occurrence frequency of the element values.

[0060] In process 902, the learning unit 108 assigns priorities among the feature quantities based on the result of process 901. When learning feature quantities in an inclusion relationship (for example, a feature quantity with a one-week period and a feature quantity with a three-week period) including the feature quantity on the inclusion side (for example, the feature quantity with a three-week period), as described as the problem to be solved by the present invention, the components of the feature quantity on the included side (for example, the feature quantity with a one-week period) will be included. Therefore, the result of the feature quantity evaluation is utilized to calculate the priority in learning.

[0061] In the inclusion relationship of periodic components, the learning unit 108 increases the priority for feature quantities with an autocorrelation coefficient equal to or greater than a threshold value, with smaller h values (i.e., smaller periods) having higher priorities. This can increase the priority of the feature quantities on the included side (such as the feature quantities with a one-week period listed in this embodiment).

[0062] Moreover, not limited to periodic ones, the feature quantities on the included side have a smaller number of unique elements compared to the feature quantities on the including side. Therefore, the learning unit 108 increases the priority of feature quantities with a smaller number of unique elements.

[0063] In addition, as a measure to make it easier for the user to understand, it is advisable to prioritize the learning of feature quantities that can represent steady fluctuations. For example, when utilizing demand forecasting in the planning operations of a logistics warehouse, it is useful for judging the validity of demand forecasting to capture the daily steady shipping fluctuations and then confirm the extent to which non-steady fluctuations have affected them. In that case, after learning about the steady fluctuation components during learning, it is necessary to learn about the non-steady fluctuations that could not be captured in that learning using a learning device capable of learning them.

[0064] In the case of feature quantities with a bias in the occurrence frequency of element values, such as fluctuations in the shipping volume caused by events that rarely occur, there is an expectation that the element values can learn local and rare fluctuations. Therefore, since such feature quantities are expected to capture non-steady fluctuations, the priority is increased for feature quantities with a small dispersion value of the occurrence frequency of element values or a large Gini coefficient value.

[0065] An example of an evaluation formula for determining the priority by comprehensively evaluating the above is the following formula (2). The smaller S is, the higher the priority. Also, k1, k2, and k3 are coefficients.

[0066]

Equation

[0067] In process 903, the learning unit 108 determines the configuration of the learning device based on prioritization. For example, when arranging the learning devices in series (boosting), the learning device at the head is applied in order from the feature with the highest priority to construct an ensemble learning device.

[0068] In process 904, the learning unit 108 performs learning using the learning device created in process 903. Here, the input of the first learning device is learned using the shipping record (i.e., demand record) of the target item, and the residual obtained by subtracting the learning fitting result from the shipping record is used as the input of the next learning device, and all subsequent learning devices are learned in the same procedure. By writing out and storing these learning results in an external file, they can be called and used in subsequent times. Figure 5 is an image of the written-out file. As the basic information of each learning device, the features used and the setting values are recorded, and if it is a serial learning device, the prediction order (prediction sequence) is recorded. In accordance with this, the information of each learned model is recorded.

[0069] In process 905, the learning unit 108 executes the prediction for the prediction period in the prediction request based on the learning model, shipping record, and prediction request. At this time, the output from each learning device becomes the contribution degree of each feature, and the sum of them becomes the prediction value.

[0070] Figure 10 is an explanatory diagram showing an example of presenting the prediction result by the demand prediction system 100 of Example 1.

[0071] The daily prediction results are drawn as shown in graph 1001 so that the time series transition can be understood. In addition to the line graph, the graph type may also be a bar graph or the like. Also, the contribution degree may be represented by a stacked bar graph 1002 with different colors or patterns for each type.

[0072] The graph 1001 illustrated in FIG. 10 is a line graph showing the prediction results for each day over a one-month period. And the graph 1002 is a stacked bar graph showing, as the basis for prediction, the contribution degree for each day, that is, the prediction results for each feature amount. For example, the prediction results based on feature amounts such as the day-of-week cycle (i.e., 7-day cycle), monthly feature amounts, and feature amounts of the end-of-month and beginning-of-month events are displayed. Thus, not only the final prediction result but also, for each prediction result, it is possible to easily grasp to what extent the prediction result based on which feature amount has an impact.

[0073] FIG. 11 is an explanatory diagram showing another presentation example of the prediction result by the demand prediction system 100 of Example 1. This is an example presented in a form different from FIG. 10.

[0074] In the example of FIG. 11, it is made possible to select for each type of contribution degree on the selection screen 1101, and it is linked with a graph 1102 that presents the prediction value when the contribution degree that has not been enabled is excluded from the prediction value. Also, the component value of the removed contribution degree may be displayed on the graph 1103. By using such notations, the influence degree of the prediction value and the contribution degree becomes clearer.

[0075] Specifically, the example of FIG. 11 is an example in which the "end-of-month and beginning-of-month event" is selected as the exclusion target on the selection screen 1101. In this case, the solid line part of the graph 1102 is the overall (i.e., including all contribution degrees) prediction result, which is the same as the graph 1001 shown in FIG. 10. In contrast, the dashed line part of the graph 1102 shows the prediction result with the contribution degree of the prediction based on the feature amount of the end-of-month and beginning-of-month event removed. And the graph 1103 shows the contribution degree of the prediction based on the removed feature amount of the end-of-month and beginning-of-month event.

[0076] Similarly, by selecting the contribution degree of the prediction based on other feature amounts as the exclusion target, its influence can be seen. Thus, it is possible to easily grasp to what extent the prediction result based on which feature amount has an impact on the prediction result.

[0077] FIG. 12 is an explanatory diagram showing an example of visualizing the evaluation process of the priorities of the feature amounts of the demand prediction system 100 according to the first embodiment.

[0078] A demand prediction user may sometimes prepare and input feature amounts by himself / herself in order to improve the prediction accuracy. At this time, when checking how the feature amounts input by the user are applied to the prediction, the display as shown in FIG. 12 is referred to. For example, the priority calculation result display 1201 may be presented, and various evaluation results in the calculation process may be presented in the form of graphs 1202 to 1204.

[0079] In the example of FIG. 12, a feature amount of a day-of-week cycle is selected. In this case, as the graph 1202, a bar graph showing the periodicity of the feature amount of the day-of-week cycle is displayed, as the graph 1203, the number of unique elements of the feature amount of the day-of-week cycle is displayed, and as the graph 1204, a histogram showing the occurrence distribution of the feature amount of the day-of-week cycle and its variance value are displayed. For example, when the feature amount generated by the user is used for prediction, the user can know how the feature amount is handled in the prediction and the basis therefor.

Embodiment

[0080] Hereinafter, a second embodiment of the present invention will be described. The demand prediction system 100 according to the second embodiment predicts the demand volume based on the warehouse shipment actual result data and the feature amount data, and presents the prediction result together with the prediction basis. In particular, it is assumed that the daily number of shipment instructions for all items is predicted and a personnel plan is made. Except for the differences described below, each part of the system according to the second embodiment has the same function as each part with the same reference numeral shown in FIGS. 1 to 12 in the first embodiment, and thus the description thereof will be omitted.

[0081] Since the picking work for the number of shipment instructions occurs, the workload is roughly proportional to the number of shipment instructions. Therefore, in a logistics warehouse, the number of personnel is often planned based on the number of shipment instructions. Therefore, the demand prediction system 100 according to the second embodiment predicts the number of shipment instructions.

[0082] This process is basically the same as in the first embodiment, but only the differences will be explained. In the process 904 in FIG. 9, if the prediction request is on a daily basis, the number of shipping instructions for all items is tallied for each day, and the tallied results are used as input to the first learning device. The other process steps are the same as in the first embodiment.

[0083] Furthermore, the system according to the embodiment of the present invention may be configured as follows.

[0084] (1) A demand forecasting system (e.g., the demand forecasting system 100) having a processor (e.g., the processor 111) and a storage device (e.g., the memory 112 or the auxiliary storage device 113), in which the storage device holds actual shipment volume (e.g., the shipment volume database 101) and a plurality of features (e.g., the feature database 103) indicating characteristics of each period of a demand volume forecast unit (e.g., each day when forecasting daily demand volume), and the processor calculates a priority of each feature volume according to information indicating a distribution of each of the plurality of feature volumes (e.g., processes 901 and 902 in FIG. 9), learns a plurality of models each of which predicts a shipment volume from each feature volume based on the priority and the actual shipment volume (e.g., processes 903 and 904 in FIG. 9), forecasts a shipment volume using the plurality of models (e.g., process 905 in FIG. 9), outputs the sum of the shipment volumes predicted using the plurality of models as a forecast value of the demand volume, and outputs the shipment volumes predicted using each model as the contribution of each feature volume to the forecast value of the demand volume (e.g., process 706 in FIG. 7 and FIG. 10).

[0085] This makes it possible to appropriately output the components of the feature in the prediction result even when there are feature values ​​that have an inclusive relationship.

[0086] (2) In (1) above, the information indicating the distribution of the features includes the order of the feature values ​​in the time direction (e.g., the order of a one-week period or a three-week period, etc.), and the processor calculates priorities based on the order of the feature values ​​in the time direction such that, among features having periodicity, features with shorter periods have a higher priority.

[0087] As a result, even when there are feature quantities in an inclusion relationship, such as a one-week cycle and a three-week cycle, the prediction result based on the included (i.e., the one with a shorter cycle) feature quantity is not included in the prediction result based on the including feature quantity, and the component of the feature quantity in the prediction result can be appropriately output.

[0088] (3) In the above (1), the information indicating the distribution of the feature quantity includes the number of types of values of the feature quantity (e.g., the number of unique elements), and the processor calculates the priority so that the feature quantity with a smaller number of types of values of the feature quantity has a higher priority.

[0089] As a result, even when there are feature quantities in an inclusion relationship, such as a one-week cycle and a three-week cycle, the prediction result based on the included (i.e., the one with a smaller number of unique elements) feature quantity is not included in the prediction result based on the including feature quantity, and the component of the feature quantity in the prediction result can be appropriately output.

[0090] (4) In the above (1), the information indicating the distribution of the feature quantity includes the bias of the distribution of the values of the feature quantity (e.g., variance value or Gini coefficient, etc.), and the processor calculates the priority so that the feature quantity with a smaller bias in the distribution of the values of the feature quantity (e.g., a smaller variance value or a larger Gini coefficient) has a higher priority.

[0091] As a result, it is possible to predict local and rare fluctuations in the element values.

[0092] (5) In the above (1), the processor learns a model for predicting the shipment quantity from the feature quantity with the highest priority, calculates the difference between the predicted value of the shipment quantity using the model learned based on the feature quantities with higher priorities and the actual performance of the shipment quantity, and further learns a model for predicting the difference from the feature quantities with lower priorities.

[0093] As a result, even when there are feature quantities in an inclusion relationship, a model for appropriately outputting the component of the feature quantity in the prediction result can be learned.

[0094] (6) In the above (1), when any one of the plurality of feature amounts is selected, the processor outputs the total of the shipment amounts predicted using a plurality of models and the total of the shipment amounts predicted using models other than the model learned based on the selected feature amount among the plurality of models (for example, FIG. 11).

[0095] Thereby, it is possible to easily grasp the influence of the prediction result based on any feature amount on the entire prediction result.

[0096] (7) In the above (1), the period of the prediction unit of the demand amount is one day, and each feature amount is a value having a predetermined period (for example, a one-week period or a three-week period, etc.) in the time direction given to each day, a value indicating the order within the month of each day, a value indicating whether each day is a holiday, a value indicating whether each day is the end of the month, a value indicating whether each day is the beginning of the month, and any one of values indicating whether each day corresponds to an arbitrary predetermined event.

[0097] Thereby, it is possible to predict the demand amount and its components for each day.

[0098] Note that the present invention is not limited to the above-described embodiments and includes various modifications. For example, the above-described embodiments have been described in detail for better understanding of the present invention and are not necessarily limited to those having all the configurations described. Also, a part of the configuration of one embodiment can be replaced with the configuration of another embodiment, and the configuration of another embodiment can be added to the configuration of one embodiment. Also, for a part of the configuration of each embodiment, addition, deletion, or replacement with other configurations is possible.

[0099] In addition, each of the above-described configurations, functions, processing units, processing means, etc. may be realized in hardware by designing a part or all of them, for example, by using an integrated circuit. Further, each of the above-described configurations, functions, etc. may be realized in software by a processor interpreting and executing a program that realizes each function. Information such as programs, tables, files, etc. that realize each function can be stored in a storage device such as a non-volatile semiconductor memory, a hard disk drive, an SSD (Solid State Drive), or a computer-readable non-transitory data storage medium such as an IC card, an SD card, or a DVD.

[0100] In addition, the control lines and information lines show those considered necessary for explanation, and not all control lines and information lines are necessarily shown on the product. In practice, it may be considered that almost all components are interconnected.

Explanation of Signs

[0101] 101 Shipment performance database 102 Prediction request database 103 Feature quantity database 104 Learning model 105 Predicted value / contribution degree database 106 Feature quantity generation unit 107 User input unit 108 Learning unit 109 Output unit 1001 Prediction result graph display screen 1002 Prediction contribution degree graph display screen 1101 Feature quantity selection screen 1102 Prediction result graph display screen 1103 Selected prediction contribution degree graph display screen 1201 Feature quantity priority display screen 1202 Feature quantity time series display screen 1203 Feature quantity evaluation 1 display screen 1204 Feature quantity evaluation 2 display screen

Claims

1. A demand prediction system having a processor and a storage device, wherein the storage device holds the actual shipment volume and a plurality of feature quantities indicating the characteristics of each period of the prediction unit of the demand volume, the plurality of feature quantities include at least a first feature quantity having a predetermined first period and a second feature quantity having a second period longer than the first period, the processor, calculates the priority of each feature quantity so that at least the priority of the first feature quantity is higher than the priority of the second feature quantity, when receiving a learning request, the processor, generates a plurality of models including a first model and a second model, inputs the actual shipment volume of each period of the past prediction unit and the first feature quantity with a higher priority among the plurality of feature quantities of each period of the prediction unit into the first model, and trains the first model to predict the shipment volume from the first feature quantity, calculates the difference between the predicted value of the shipment volume of each period of the prediction unit using the first model and the actual shipment volume, and inputs the difference and the second feature quantity with a lower priority among the plurality of feature quantities of each period of the prediction unit into the second model, and trains the second model to predict the difference from the second feature quantity, when receiving a prediction request, the processor, inputs the first feature quantity of each period of the prediction unit to be predicted into the first model to predict the shipment volume of each period of the prediction unit, inputs the second feature quantity of each period of the prediction unit to be predicted into the second model to predict the difference of each period of the prediction unit, outputs, as the predicted value of the demand volume of each period of the prediction unit, the value obtained by summing up the shipment volume and the difference predicted using the plurality of models for each period of the prediction unit, and outputs, as the contribution degree of each feature quantity to the predicted value of the demand volume of each period of the prediction unit, the shipment volume and the difference of each period of the prediction unit predicted using each model. A demand prediction system characterized by this.

2. The demand prediction system according to claim 1, wherein the processor calculates the priority so that, among the plurality of feature quantities, the feature quantity with a smaller number of value types has a higher priority. A demand prediction system characterized by this.

3. The demand prediction system according to claim 1, The demand prediction system is characterized in that the processor calculates the priority such that, among the plurality of feature quantities, a feature quantity with a smaller bias in the distribution of values has a higher priority.

4. The demand prediction system according to claim 1, wherein the processor outputs a value obtained by summing the shipment quantities and the differences predicted using the plurality of models for each period of the prediction unit, and when any one of the plurality of feature quantities is selected, the shipment quantity or the difference predicted using a model learned based on the selected feature quantity is excluded from the value obtained by summing the shipment quantities and the differences predicted using the plurality of models for each period of the prediction unit, and further outputs the resulting value.

5. The demand prediction system according to claim 1, wherein the period of the prediction unit of the demand quantity is one day, and each of the feature quantities is any one of a value having a predetermined period in the time direction assigned to each day, a value indicating the order within the month for each day, a value indicating whether each day is a holiday, a value indicating whether each day is the end of the month, a value indicating whether each day is the beginning of the month, and a value indicating whether each day corresponds to an arbitrary predetermined event.

6. A demand prediction method executed by a computer system having a processor and a storage device, wherein the storage device holds the actual shipment quantity and a plurality of feature quantities indicating the features of each period of the prediction unit of the demand quantity, the plurality of feature quantities including at least a first feature quantity having a predetermined first period and a second feature quantity having a second period longer than the first period, and the demand prediction method includes a first step in which the processor calculates the priority of each feature quantity such that at least the priority of the first feature quantity is higher than the priority of the second feature quantity, When receiving a learning request, the processor generates a plurality of models including a first model and a second model, inputs the actual shipment volume of each period of the past prediction unit and the first feature quantity with a high priority among the plurality of feature quantities of each period of the prediction unit into the first model, and trains the first model to predict the shipment volume from the first feature quantity. Calculate the difference between the predicted value of the shipment volume of each period of the prediction unit using the first model and the actual performance of the shipment volume of each period of the prediction unit, and input the difference and the second feature quantity with a low priority among the plurality of feature quantities of each period of the prediction unit into the second model, and train the second model to predict the difference from the second feature quantity; a second procedure; When receiving a prediction request, the processor inputs the first feature quantity of each period of the prediction unit to be predicted into the first model to predict the shipment volume of each period of the prediction unit, and inputs the second feature quantity of each period of the prediction unit to be predicted into the second model to predict the difference of each period of the prediction unit; a third procedure; The processor outputs, as the predicted value of the demand volume of each period of the prediction unit, the value obtained by summing up the shipment volume and the difference predicted using the plurality of models for each period of the prediction unit, and outputs, as the contribution degree of each feature quantity to the predicted value of the demand volume of each period of the prediction unit, the shipment volume and the difference of each period of the prediction unit predicted using each model; a fourth procedure, characterized in that the demand prediction method includes the above.

7. The demand prediction method according to claim 6, In the first procedure, the processor calculates the priority such that, among the plurality of feature quantities, the feature quantity with a smaller number of value types has a higher priority. The demand prediction method is characterized by this.

8. The demand prediction method according to claim 6, In the first procedure, the processor calculates the priority such that, among the plurality of feature quantities, the feature quantity with a smaller deviation in the value distribution has a higher priority. The demand prediction method is characterized by this.

9. The demand prediction method according to claim 6, In the fourth procedure, the processor outputs a value obtained by summing up the shipment quantity predicted using the plurality of models and the difference for each period of the prediction unit. When any one of the plurality of feature amounts is selected, a value obtained by excluding the shipment quantity or the difference predicted using the model learned based on the selected feature amount from the value obtained by summing up the shipment quantity predicted using the plurality of models and the difference for each period of the prediction unit is further output. A demand prediction method characterized by the above.

10. A demand prediction method according to claim 6, wherein the period of the prediction unit of the demand quantity is one day, the plurality of feature amounts include at least two of a value having a predetermined period in the time direction given to each day, a value indicating the order within the month of each day, a value indicating whether each day is a holiday, a value indicating whether each day is the end of the month, a value indicating whether each day is the beginning of the month, and a value indicating whether each day corresponds to an arbitrary predetermined event. A demand prediction method characterized by the above.

Images