Demand forecasting system, deman forecasting method, and program

Abstract

To provide a demand forecasting system capable of performing highly accurate demand forecasting while reducing system construction time and cost.SOLUTION: A demand forecasting system comprises: data acquisition means for acquiring time-series demand data; data extraction means for extracting trend components and residual components from the demand data; normalization means for normalizing the residual components on the basis of the maximum and minimum elements of the residual components; moment information calculation means for calculating moment information at each time point on the basis of the normalized residual components; machine learning means for training a regression model where the moment information serves as input information and the residual components serve as output information; trend component forecasting means for calculating a future trend component forecasting value using an equation of a linear regression model approximated by a polynomial in time applied to the trend component; and demand forecasting means for calculating a demand forecasting value on the basis of the future residual component forecasting value output from the trained regression model of the machine learning means and the trend component forecasting value.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present disclosure relates to a demand forecasting system, a demand forecasting method, and a program for forecasting demand. [Background technology]

[0002] A prediction system that uses a prediction model generated by machine learning of big data, etc., to accurately predict complex demand characteristics is known (see, for example, Patent Document 1). This prediction system performs highly accurate demand prediction by classifying the patterns of meta information in big data and constructing a prediction model for each pattern. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2023-100466 A Summary of the Invention [Problem to be solved by the invention]

[0004] However, in the above prediction system, due to constraints on resources and data management costs, it is not easy to build a database of such big data, and there is a risk that building the system will take time and cost a lot.

[0005] The present disclosure has been made to solve such problems, and has as its primary objective the provision of a demand forecasting system, a demand forecasting method, and a program that can perform highly accurate demand forecasts while reducing the time and cost of system construction. [Means for solving the problem]

[0006] In order to achieve the above object, one aspect of the present disclosure is to A data acquisition means for acquiring time-series demand data; a data extraction means for extracting a trend component and a residual component of the demand data from the demand data acquired by the data acquisition means; a normalization means for identifying a maximum element and a minimum element from among the residual components extracted by the data extraction means, and normalizing the residual components based on the identified maximum element and minimum element; a moment information calculation means for calculating moment information at each time based on the residual component normalized by the normalization means; a machine learning means for learning a regression model using the moment information as input information and the residual component as output information; a trend component prediction means for creating an equation of a linear regression model that approximates the trend component extracted by the data extraction means with a polynomial related to time, and calculating a trend component predicted value that is a predicted value of the future trend component using the created equation; a demand forecasting means for calculating a demand forecast value based on a residual component forecast value, which is a forecast value of a future residual component output from the regression model of the machine learning means after learning, and a trend component forecast value calculated by the trend component forecasting means; This is a demand forecasting system. In this aspect, the moment information calculation means may calculate high-order moment information having a moment order of 5 or more. In order to achieve the above object, one aspect of the present disclosure is to obtaining time series demand data; extracting a trend component and a residual component of the demand data from the acquired demand data; identifying a maximum element and a minimum element from among the extracted residual components, and normalizing the residual components based on the identified maximum element and minimum element; calculating moment information at each time based on the normalized residual components; learning a regression model using the moment information as input information and the residual component as output information; A step of creating an equation of a linear regression model that approximates the extracted trend component with a polynomial related to time, and calculating a trend component forecast value that is a forecast value of the future trend component using the created equation; calculating a demand forecast value based on a residual component forecast value, which is a forecast value of a future residual component output from the regression model of the machine learning means after the learning, and the calculated trend component forecast value; The demand forecasting method includes: In order to achieve the above object, one aspect of the present disclosure is to A process of acquiring time series demand data; A process of extracting a trend component and a residual component of the demand data from the acquired demand data; a process of identifying a maximum element and a minimum element from among the extracted residual components, and normalizing the residual components based on the identified maximum element and minimum element; A process of calculating moment information at each time based on the normalized residual components; a process of learning a regression model in which the moment information is input information and the residual component is output information; A process of creating an equation of a linear regression model that approximates the extracted trend component with a polynomial related to time, and calculating a trend component forecast value that is a forecast value of the future trend component using the created equation; A process of calculating a demand forecast value based on a residual component forecast value, which is a forecast value of a future residual component output from the regression model of the machine learning means after the learning, and the calculated trend component forecast value; It is a program that causes a computer to execute the above. Effect of the Invention

[0007] According to the present disclosure, it is possible to provide a demand forecasting system, a demand forecasting method, and a program that can perform highly accurate demand forecasting while reducing the time and cost of system construction. [Brief description of the drawings]

[0008] [Figure 1] 1 is a block diagram showing a schematic system configuration of a demand forecasting system according to an embodiment of the present invention. [Diagram 2] 2 is a flowchart showing a flow of a demand forecasting method according to the present embodiment. [Diagram 3] FIG. 11 is a diagram showing a simulation result by the demand forecasting system. [Figure 4] FIG. 11 is a diagram showing a simulation result by the demand forecasting system. [Diagram 5] FIG. 11 is a diagram showing a simulation result by the demand forecasting system. [Figure 6] FIG. 11 is a diagram showing a simulation result by the demand forecasting system. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] Hereinafter, the present embodiment will be described with reference to the drawings. For example, in logistics for supplying service parts (SP) to a production line, production preparation and inventory adjustment are performed based on the demand forecast results in order to reduce shortages and excess inventory of service parts. For this reason, highly accurate demand forecasting is desired. On the other hand, demand patterns are diverse and include not only trend components due to aging but also periodic fluctuation components and irregular noise components due to seasonal influences, and various demand characteristics are mixed. In addition to such complex demand characteristics, under the constraints of resources and man-hours, there is a constraint to keep demand performance data and related information (meta information, etc.) to a minimum. For this reason, there is a risk that the accuracy of demand forecasting will decrease.

[0010] In contrast, the demand forecasting system according to this embodiment extracts residual components other than trend components from the time-series distribution of actual demand data, calculates higher-order moment information of the residual components, and performs machine learning using the higher-order moment information as features to forecast demand.

[0011] As a result, for example, by effectively utilizing high-order moment information that can be calculated even under restricted or scarce demand performance data and meta-information, it is possible to reduce the number of feature quantities required for constructing a prediction model while performing highly accurate demand prediction. That is, it is possible to perform highly accurate demand prediction while reducing the time and cost of system construction. Furthermore, it may lead to an expansion of the application range of demand prediction by machine learning.

[0012] Generally, when performing demand prediction by machine learning using combinations of demand performance data for N months, N more than two feature quantity data are desired. For example, when N = 12, 12 = more than 4096 feature quantity data are required.

[0013] On the other hand, according to the present embodiment, the feature representation can be reduced to a combination of K (K < N) high-order moment components, and the necessary feature quantity data can be reduced to K two. For example, considering high-order moments up to K = 5, 5 = 32 feature quantity data are sufficient, so the number of required feature quantity data can be significantly reduced.

[0014] FIG. 1 is a block diagram showing a schematic system configuration of a demand prediction system according to the present embodiment. The demand prediction system 1 according to the present embodiment includes a data acquisition unit 2, a data extraction unit 3, a normalization unit 4, a moment information calculation unit 5, a machine learning unit 6, a trend component prediction unit 7, and a demand prediction unit 8.

[0015] The demand forecasting system 1 has a typical computer hardware configuration including, for example, a processor such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit), internal memory such as a RAM (Random Access Memory) or a ROM (Read Only Memory), storage devices such as an HDD (Hard Disk Drive) or an SDD (Solid State Drive), an input / output I / F for connecting peripheral devices such as a display, and a communication I / F for communicating with devices outside the device.

[0016] The data acquisition unit 2 is a specific example of a data acquisition means. The data acquisition unit 2 acquires demand data indicating a time-series demand record. The demand data may be input to the data acquisition unit 2 by a user using an input device, for example.

[0017] The demand data is, for example, order record data Y for each part number of service parts (SP). The order record data Y is time series data {y1, y2, . . . , y t} and y i (1≦i≦t) may be the order record data at time i in the past.

[0018] The data extraction unit 3 is a specific example of a data extraction means. The data extraction unit 3 decomposes the order performance data Y acquired by the data acquisition unit 2 into a trend component T indicating a trend of the order performance data and a remaining residual component R, and extracts the trend component T and the residual component R of the order performance data Y, respectively. Y=T+R Here, the trend component T = {T1, T2, . . . , T t}, and the residual component R = {R1, R2, . . . , R t}.

[0019] The normalization unit 4 is a specific example of a normalization means. The normalization unit 4 normalizes the residual components R={R1, R2, . . . , R t}, the maximum element R maxand the minimum element R min And, identify.

[0020] Then, the normalization unit 4 calculates the following formula and the specified maximum element R max and the smallest element R min Based on this, the residual component R is normalized so that its maximum value is 1 and its minimum value is 0, and the normalized residual component {R ’ i} i=1,t Hereinafter, normalized parameters will be marked with a dash (') in the upper right corner.

number

[0021] The moment information calculation unit 5 is a specific example of a moment information calculation means. The moment information calculation unit 5 calculates the residual component {R ’ i} i=1,t Based on this, the moment information {M k i} k=1,K i=N,t Calculate.

number

[0022] In the above formula, k represents the order of the moment, and K represents the maximum order of the moment required for prediction. For example, if the number of pieces of order record data for individual part numbers, that is, the number of pieces of time-series data, is D, then K is 2. K K is set so that .about.D.

[0023] In this embodiment, for example, the moment information calculation unit 5 calculates the higher-order moment information with K=5, but K may be 6 or more. The number of most recent order data required to represent the moment information at each time i is defined as N.

[0024] The machine learning unit 6 is a specific example of a machine learning means. The machine learning unit 6 calculates the moment information {M k i} k=1,K i=N+1,t-n is the input information, and the residual component R ’ i+n Here, N+1≦i≦tn, and the machine learning algorithm of the regression model is, for example, SVR (Support Vector Regressor) or GBR (Gradient Boosting Regressor).

[0025] The regression model of the machine learning unit 6 after the above learning is the moment information {M k i} k=1,K When is input, the future residual component prediction value (hereinafter, residual component prediction value) R ’- i+n Hereinafter, the predicted value will be indicated by adding a bar (-) to the upper right of the parameter. Also, t-n+1≦i≦t.

[0026] The trend component prediction unit 7 is a specific example of a trend component prediction means. The trend component prediction unit 7 predicts the trend components T={T1, T2, . . . , T t For example, the following linear regression model equation is created by approximating the trend component T with a third-order polynomial for time i (1≦i≦t). Note that the trend component T monotonically increases or decreases. For this reason, in this embodiment, approximation is performed with a third-order polynomial as described above, but the present invention is not limited to this and may be approximated with, for example, a fourth-order or higher polynomial.

number

[0027] The trend component prediction unit 7 uses the equation of the linear regression model created above to calculate a trend component prediction value T - i+n In the above formula, t-n+1≦i≦t.

[0028] The demand forecasting unit 8 is a specific example of a demand forecasting means. The demand forecasting unit 8 calculates the residual component predicted value R ’- i+n and the maximum element R identified by the normalization unit 4 max and the smallest element R min Based on this, the residual component prediction value R before normalization is calculated using the following formula: - i+n Calculate.

number

[0029] Furthermore, the demand forecasting unit 8 calculates the residual component forecast value R - i+n and the trend component predicted value T calculated by the trend component prediction unit 7. - i+n Based on this, the demand forecast value y - i+n Calculate.

number

[0030] Next, the demand forecasting method according to this embodiment will be described. Fig. 2 is a flowchart showing the flow of the demand forecasting method according to this embodiment. The data acquisition unit 2 acquires time-series demand data Y (step S101).

[0031] The data extraction unit 3 decomposes the demand data Y acquired by the data acquisition unit 2 into a trend component T and a residual component R, and extracts the trend component T and the residual component R of the demand data (step S102).

[0032] The normalization unit 4 extracts the maximum element R from the residual components R extracted by the data extraction unit 3. max and the smallest element R min Identify the maximum element R max and the smallest element R minThe residual component R is normalized based on the above (step S103).

[0033] The moment information calculation unit 5 calculates the residual component {R ’ i} i=1,t Based on this, the moment information {M k i} k=1,K i=N,t is calculated (step S104).

[0034] The machine learning unit 6 calculates the moment information {M k i} k=1,K i=N,t is the input information, and the residual component R ’ i+n A regression model with the above as output information is trained (step S105).

[0035] The trend component prediction unit 7 creates a linear regression model equation that approximates the trend component T extracted by the data extraction unit 3 with, for example, a third-order polynomial with respect to time i, and calculates a trend component predicted value T - i+n is calculated (step S106).

[0036] The demand forecasting unit 8 calculates the residual component predicted value R ’- i+n and the maximum element R identified by the normalization unit 4 max and the smallest element R min Based on this, the residual component prediction value R - i+n The demand forecasting unit 8 calculates the residual component predicted value R - i+n and the trend component predicted value T calculated by the trend component prediction unit 7. - i+n Based on this, the demand forecast value y - i+n is calculated (step S107).

[0037] Next, a simulation result of comparing demand forecasts made by the demand forecasting system 1 according to the present embodiment and a conventional demand forecasting system will be described. Fig. 3 and Fig. 4 are diagrams showing the simulation results made by the demand forecasting system.

[0038] 3 and 4, the upper part (a) shows the simulation results using a conventional demand forecasting system, and the lower part (b) shows the simulation results using the demand forecasting system 1 according to this embodiment.

[0039] In the simulation, actual order data for automobile service parts is used as the demand data. In Figures 3 and 4, the dotted lines show the actual order data for each representative part over a 36-month period, and the solid lines show the demand forecast values ​​for the most recent year, which are predicted by each demand forecasting system based on the actual order data.

[0040] As shown in the upper part (a) of Figures 3 and 4, it can be seen that in the conventional demand forecasting system, the demand forecast value (solid line) deviates significantly from the actual order receipt data (dotted line). On the other hand, as shown in the lower part (b) of Figures 3 and 4, in the demand forecasting system 1 according to this embodiment, the actual order receipt data (dotted line) and the demand forecast value (solid line) approximately match, and it can be seen that demand can be forecast with higher accuracy.

[0041] Furthermore, the conventional demand forecasting system performs demand forecasting after clustering similar features between parts and extracting feature amounts using a database of order records of approximately 9,000 parts, and therefore requires a large amount of demand data. On the other hand, the demand forecasting system 1 according to the present embodiment performs demand forecasting using only order record data for each part, i.e., single time-series data, and therefore it can be confirmed that more accurate demand forecasting is possible with less demand data, as described above.

[0042] Next, another simulation result in which demand forecasting is performed using the demand forecasting system 1 according to this embodiment and a conventional demand forecasting system and the results are compared will be described. Figures 5 and 6 are diagrams showing the simulation results using the demand forecasting system. The upper parts (a) of Figures 5 and 6 show the simulation results using the conventional demand forecasting system, and the lower parts (b) of Figures 5 and 6 show the simulation results using the demand forecasting system 1 according to this embodiment.

[0043] In this simulation, the demand forecasting system 1 according to this embodiment sets the order k of the moment to 5 (k=5) and performs the above-mentioned demand forecast. On the other hand, the conventional demand forecasting system sets the order k of the moment to 1 (k=1) and performs demand forecasting using a moving average forecasting method. In this moving average forecasting method, the demand forecast value is the average value of the actual orders received in the most recent 12 months at the time of the forecast.

[0044] As shown in the upper part (a) of Fig. 5 and Fig. 6, in the conventional demand forecasting system, it can be seen that the demand forecast value (solid line) largely deviates from the actual order receipt data (dotted line). On the other hand, as shown in the lower part (b) of Fig. 5 and Fig. 6, in the demand forecasting system 1 according to the present embodiment, the actual order receipt data (dotted line) and the demand forecast value (solid line) roughly match, and it can be seen that the demand can be forecast with higher accuracy.

[0045] From the above simulation results, it can be confirmed that the demand forecasting system 1 according to this embodiment can accurately forecast various demands even under conditions where the number of demand data is small. Therefore, the effectiveness of the demand forecasting system 1 according to this embodiment can be confirmed.

[0046] Although some embodiments of the present disclosure have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included in the scope and spirit of the invention, and are included in the scope of the invention and its equivalents described in the claims.

[0047] The present disclosure can also be implemented by causing a processor to execute a computer program to perform the processes shown in FIG.

[0048] The program can be stored and provided to a computer using various types of non-transitory computer readable media. The non-transitory computer readable media includes various types of tangible storage media. Examples of the non-transitory computer readable media include magnetic recording media (e.g., flexible disks, magnetic tapes, hard disk drives), magneto-optical recording media (e.g., magneto-optical disks), CD-ROMs (Read Only Memory), CD-Rs, CD-R / Ws, and semiconductor memories (e.g., mask ROMs, PROMs (Programmable ROMs), EPROMs (Erasable PROMs), flash ROMs, and RAMs (random access memories)).

[0049] The program may be provided to the computer by various types of transitory computer readable media. Examples of the transitory computer readable medium include an electric signal, an optical signal, and an electromagnetic wave. The transitory computer readable medium can provide the program to the computer via a wired communication path such as an electric wire and an optical fiber, or via a wireless communication path.

[0050] Each component of the demand forecasting system 1 according to the above-described embodiment can be realized not only by a program, but also in whole or in part by dedicated hardware such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field-Programmable Gate Array). [Explanation of symbols]

[0051] 1 Demand forecasting system, 2 Data acquisition section, 3 Data extraction section, 4 Normalization section, 5 Moment information calculation section, 6 Machine learning section, 7 Trend component forecasting section, 8 Demand forecasting section

Claims

1. A data acquisition means for acquiring time-series demand data; a data extraction means for extracting a trend component and a residual component of the demand data from the demand data acquired by the data acquisition means; a normalization means for identifying a maximum element and a minimum element from among the residual components extracted by the data extraction means, and normalizing the residual components based on the identified maximum element and minimum element; a moment information calculation means for calculating moment information at each time based on the residual component normalized by the normalization means; a machine learning means for learning a regression model using the moment information as input information and the residual component as output information; a trend component prediction means for creating an equation of a linear regression model that approximates the trend component extracted by the data extraction means with a polynomial related to time, and calculating a trend component predicted value that is a predicted value of the future trend component using the created equation; a demand forecasting means for calculating a demand forecast value based on a residual component forecast value, which is a forecast value of a future residual component output from the regression model of the machine learning means after learning, and a trend component forecast value calculated by the trend component forecasting means; A demand forecasting system.

2. 2. The demand forecasting system according to claim 1, The moment information calculation means calculates high-order moment information having a moment order of 5 or more.

3. obtaining time series demand data; extracting a trend component and a residual component of the demand data from the acquired demand data; identifying a maximum element and a minimum element from among the extracted residual components, and normalizing the residual components based on the identified maximum element and minimum element; calculating moment information at each time based on the normalized residual components; learning a regression model using the moment information as input information and the residual component as output information; A step of creating an equation of a linear regression model that approximates the extracted trend component with a polynomial related to time, and calculating a trend component forecast value that is a forecast value of the future trend component using the created equation; calculating a demand forecast value based on a residual component forecast value, which is a forecast value of a future residual component output from the regression model of the machine learning means after the learning, and the calculated trend component forecast value; A demand forecasting method, including:

4. A process of acquiring time series demand data; A process of extracting a trend component and a residual component of the demand data from the acquired demand data; a process of identifying a maximum element and a minimum element from among the extracted residual components, and normalizing the residual components based on the identified maximum element and minimum element; A process of calculating moment information at each time based on the normalized residual components; a process of learning a regression model in which the moment information is input information and the residual component is output information; A process of creating an equation of a linear regression model that approximates the extracted trend component with a polynomial related to time, and calculating a trend component forecast value that is a forecast value of the future trend component using the created equation; A process of calculating a demand forecast value based on a residual component forecast value, which is a forecast value of a future residual component output from the regression model of the machine learning means after the learning, and the calculated trend component forecast value; A program that causes a computer to execute the following.

Images