Demand forecasting system, demand forecasting method, and program

Abstract

To perform forecasting with sufficient precision also in consideration of variations in different time series scales with respect to a plurality of types of provision targets.SOLUTION: A demand forecasting system inputs actual result data including the number of time series actual demand results for each of a plurality of types of provision targets, executes a time series clustering process, extracts trend component data for each cluster, calculates pattern data by performing a statistical process on remaining data for a predetermined period, and obtains first forecasting data indicating a representative demand forecasting value of the provision target for each cluster by calculating a sum of the pattern data and the trend component data for each cluster. The demand forecasting system inputs, for each provision target, a difference between the actual result data and the first forecasting data in a learning model in which a regression learning is performed on the difference, obtains second forecasting data that is data indicating a forecasting value of the difference for each provision target, and obtains data indicating a demand forecasting value for each provision target by calculating a sum of the first forecasting data and the second forecasting data for each provision target.SELECTED DRAWING: Figure 1

Description

[Technical field]

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

[0002] Patent Document 1 describes a prediction system that makes predictions based on past time-series data. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2021-57010 Summary of the Invention [Problem to be solved by the invention]

[0004] In the technology described in Patent Document 1, clustering is performed on smoothed performance data and a predictor is created for each cluster, so information on short-term fluctuations in the time series data is missing at the time of prediction, making it impossible to perform highly accurate predictions. In addition, the technology described in Patent Document 1 does not anticipate clustering multiple time series with different scales.

[0005] The present disclosure has been made to solve such problems, and its purpose is to provide a demand forecasting system, a demand forecasting method, and a program that are capable of predicting demand for goods or services, which are provided, with sufficient accuracy, taking into account fluctuations on different time series scales for multiple types of provided objects, as well as to provide a learning system, a learning method, and a program that are capable of generating a learning model to be used for such predictions. [Means for solving the problem]

[0006] A demand forecasting system according to a first aspect of the present disclosure inputs, for provision targets which are goods or services, performance data including a time-series actual number of demands for each of a plurality of types of provision targets, executes a time-series clustering process on the input performance data for each of the provision targets, executes an extraction process for each cluster classified by the time-series clustering process to extract trend component data for the cluster from the performance data, and executes a pattern calculation process for calculating pattern data indicative of a demand pattern for the predetermined period by performing a statistical process on first data, which is data remaining after the trend component data is extracted from the performance data for each cluster, a first prediction process for obtaining, for each cluster, first prediction data indicating a representative demand prediction value of the provision object belonging to the cluster by calculating a sum of the pattern data and the trend component data for each data, a second prediction process for obtaining, for each provision object, second prediction data which is data indicating a prediction value of the difference for each provision object by inputting a difference between the performance data of the provision object and the first prediction data for the cluster to which the provision object belongs to a learning model which has been regression-trained on the difference for the provision object, and a third prediction data which is data indicating a demand prediction value for each provision object by calculating a sum of the first prediction data and the second prediction data for each provision object. With such a configuration, the demand prediction system can predict the demand for the provision objects which are goods or services with sufficient accuracy, taking into consideration fluctuations on different time series scales for a plurality of types of provision objects.

[0007] The demand forecasting system may execute a determination process for determining whether or not a trend component is contained for each cluster on post-extraction data, which is data after data of the trend component is extracted in the extraction process, and when it is determined as a result of the determination process that a trend component is contained for at least one of the clusters, execute the time series clustering process and the extraction process on the post-extraction data instead of the actual data until it is determined in the determination process that the post-extraction data does not contain a trend component, and the pattern calculation process may extract data of the trend component from the post-extraction data instead of the actual data at a stage when it is determined in the determination process that the data does not contain a trend component, and calculate the pattern data by performing a statistical process on the first data for the predetermined period.

[0008] The specified period may be a period whose start and end dates are expressed in terms of days or months in a year, and the statistical processing may be a process of calculating statistical values ​​of the first data for the specified period in the first data for a specified number of years, for each cluster.

[0009] The first prediction process may include an approximation process of approximating the data of the trend component extracted in the extraction process to a polynomial, and may obtain the first prediction data for each cluster by calculating, for each cluster, a sum of the pattern data and data represented by the polynomial.

[0010] In the time series clustering process, the extraction process, and the pattern calculation process, standardized data for the multiple types of objects to be provided may be used as the performance data, and in the second prediction process, data before the standardization is performed may be used as the performance data.

[0011] The item to be provided may be any one of parts constituting one or more products, a predetermined amount of materials used in manufacturing one or more products, a product belonging to one or more product groups, and a service belonging to one or more service groups.

[0012] A learning system according to a second aspect of the present disclosure inputs into a regression learning model a difference between actual data including a time-series actual demand count for each of a plurality of types of provision targets, the actual data being for each of the provision targets, which are goods or services, and first prediction data, which is data indicating a representative demand prediction value for the cluster to which the provision targets belong, obtained for each cluster classified by performing a time-series clustering process on the actual data for each of the provision targets, and executes regression learning of the regression learning model. The regression learning model is updated during operation to become a model that inputs the difference and outputs a prediction value of the difference for a predetermined period. The first prediction data is obtained as data indicating a representative demand prediction value of the provision targets belonging to the cluster for each cluster by executing, for each cluster, an extraction process for extracting trend component data of the cluster from the actual data, and for each cluster, a pattern calculation process for calculating pattern data indicating a demand pattern for the predetermined period by performing a statistical process on first data, which is data remaining after extracting the trend component data from the actual data, and a first prediction process for calculating a sum of the pattern data and the trend component data, thereby obtaining data indicating a representative demand prediction value of the provision targets belonging to the cluster for each cluster. With this configuration, the learning system is capable of generating a learning model to be used in a demand forecasting system that can predict demand for goods or services with sufficient accuracy, taking into account fluctuations on different time series scales for multiple types of goods or services.

[0013] The specified period may be a period whose start and end dates are expressed in terms of days or months in a year, and the statistical processing may be a process of calculating statistical values ​​of the first data for the specified period in the first data for a specified number of years, for each cluster.

[0014] In the time series clustering process, the extraction process, and the pattern calculation process, standardized data for the multiple types of objects to be provided may be used as the performance data, and the difference input to the regression learning model may be a difference calculated using data before the standardization as the performance data.

[0015] The item to be provided may be any one of parts constituting one or more products, a predetermined amount of materials used in manufacturing one or more products, a product belonging to one or more product groups, and a service belonging to one or more service groups.

[0016] A demand forecasting method according to a third aspect of the present disclosure includes inputting, for provision targets which are goods or services, performance data including a time-series actual demand count for each of a plurality of types of provision targets, executing a time-series clustering process on the input performance data for each of the provision targets, executing an extraction process for extracting trend component data for the cluster from the performance data for each cluster classified by the time-series clustering process, and executing a pattern calculation process for calculating pattern data indicative of a demand pattern for the predetermined period by performing statistical processing on first data, which is data remaining after the trend component data is extracted from the performance data for each cluster, a first prediction process for obtaining, for each cluster, first prediction data indicating a representative demand prediction value of the provision object belonging to the cluster by calculating a sum of the pattern data and the trend component data, a second prediction process for obtaining, for each provision object, second prediction data which is data indicating a prediction value of the difference for the provision object by inputting a difference between the actual data of the provision object and the first prediction data for the cluster to which the provision object belongs to a learning model which has been regression-trained on the difference for the provision object, and a third prediction data which is data indicating a demand prediction value for the provision object by calculating a sum of the first prediction data and the second prediction data for each provision object. With this configuration, the demand prediction method makes it possible to predict the demand for the provision objects, which are goods or services, with sufficient accuracy, taking into account fluctuations on different time series scales for a plurality of types of provision objects.

[0017] A determination process is executed on extracted data, which is data after the data of the trend component is extracted in the extraction process, to determine whether or not the data contains a trend component for each cluster. When it is determined as a result of the determination process that at least one of the clusters contains a trend component, the time series clustering process and the extraction process are executed on the extracted data instead of the performance data until it is determined in the determination process that the extracted data does not contain a trend component. The pattern calculation process may be configured to extract the data of the trend component from the extracted data instead of the performance data at a stage when it is determined in the determination process that the data does not contain a trend component, and to use the remaining data as the first data and to calculate the pattern data by performing a statistical process on the first data for the predetermined period.

[0018] The specified period may be a period whose start and end dates are expressed in terms of days or months in a year, and the statistical processing may be a process of calculating statistical values ​​of the first data for the specified period in the first data for a specified number of years, for each cluster.

[0019] The first prediction process may include an approximation process of approximating the data of the trend component extracted in the extraction process to a polynomial, and may obtain the first prediction data for each cluster by calculating, for each cluster, a sum of the pattern data and data represented by the polynomial.

[0020] In the time series clustering process, the extraction process, and the pattern calculation process, standardized data for the multiple types of objects to be provided may be used as the performance data, and in the second prediction process, data before the standardization is performed may be used as the performance data.

[0021] The item to be provided may be any one of parts constituting one or more products, a predetermined amount of materials used in manufacturing one or more products, a product belonging to one or more product groups, and a service belonging to one or more service groups.

[0022] A learning method according to a fourth aspect of the present disclosure includes inputting into a regression learning model a difference between actual data including a time-series actual demand count for each of a plurality of types of provision targets, the actual data being for each of the provision targets, which are goods or services, and first prediction data, which is data indicating a representative demand prediction value for the cluster to which the provision targets belong, obtained for each cluster classified by performing a time-series clustering process on the actual data for each of the provision targets, and executing regression learning of the regression learning model. The regression learning model is updated so that, during operation, the regression learning model becomes a model that inputs the difference and outputs a prediction value of the difference for a predetermined period. The first prediction data is obtained as data indicating a representative demand prediction value of the provision targets belonging to the cluster for each cluster by performing an extraction process for each cluster to extract trend component data of the cluster from the actual data, performing a pattern calculation process for calculating pattern data indicating a demand pattern for the predetermined period by performing a statistical process for each cluster on first data, which is data remaining after extracting the trend component data from the actual data, and performing a first prediction process for each cluster to calculate a sum of the pattern data and the trend component data. With such a configuration, the learning method makes it possible to generate a learning model to be used in a demand forecasting system that can predict demand for goods or services with sufficient accuracy, taking into account fluctuations on different time series scales for multiple types of goods or services.

[0023] The specified period may be a period whose start and end dates are expressed in terms of days or months in a year, and the statistical processing may be a process of calculating statistical values ​​of the first data for the specified period in the first data for a specified number of years, for each cluster.

[0024] In the time series clustering process, the extraction process, and the pattern calculation process, standardized data for the multiple types of objects to be provided may be used as the performance data, and the difference input to the regression learning model may be a difference calculated using data before the standardization as the performance data.

[0025] The item to be provided may be any one of parts constituting one or more products, a predetermined amount of materials used in manufacturing one or more products, a product belonging to one or more product groups, and a service belonging to one or more service groups.

[0026] A program according to a fifth aspect of the present disclosure includes inputting, into a computer, performance data including a time-series actual number of demands for each of a plurality of types of provision targets which are goods or services, executing a time-series clustering process on the performance data for each of the input provision targets, executing an extraction process for extracting trend component data for the cluster from the performance data for each cluster classified by the time-series clustering process, and executing a pattern calculation process for calculating pattern data indicating a demand pattern for the predetermined period by performing statistical processing on first data which is data remaining after extracting the trend component data from the performance data for each cluster, A program for executing a demand forecasting process, the program comprising: a first forecasting process for obtaining, for each cluster, first forecasting data indicating a representative demand forecast value of the object of supply belonging to the cluster by calculating a sum of turn data and data of the trend component; a second forecasting process for obtaining, for each object of supply, second forecasting data which is data indicating a forecast value of the difference between the performance data of the object of supply and the first forecasting data for the cluster to which the object of supply belongs, inputting the difference between the performance data of the object of supply and the first forecasting data for the cluster to which the object of supply belongs to a learning model which has been regression-trained on the difference for the object of supply, and a third forecasting data which is data indicating a demand forecast value for the object of supply by calculating a sum of the first forecasting data and the second forecasting data for each object of supply. With such a configuration, the program makes it possible to predict with sufficient accuracy the demand for the object of supply, which is a product or a service, taking into consideration fluctuations on different time series scales for a plurality of types of objects of supply.

[0027] The demand forecasting process may execute a determination process for determining whether or not a trend component is contained for each cluster on extracted data, which is data after data of the trend component is extracted in the extraction process, and when it is determined as a result of the determination process that a trend component is contained for at least one of the clusters, execute the time series clustering process and the extraction process on the extracted data instead of the actual data until it is determined in the determination process that the extracted data does not contain a trend component, and the pattern calculation process may extract data of the trend component from the extracted data instead of the actual data at a stage when it is determined in the determination process that the data does not contain a trend component, and use the remaining data as the first data and calculate the pattern data by performing statistical processing on the first data for the predetermined period.

[0028] The specified period may be a period whose start and end dates are expressed in terms of days or months in a year, and the statistical processing may be a process of calculating statistical values ​​of the first data for the specified period in the first data for a specified number of years, for each cluster.

[0029] The first prediction process may include an approximation process of approximating the data of the trend component extracted in the extraction process to a polynomial, and may obtain the first prediction data for each cluster by calculating, for each cluster, a sum of the pattern data and data represented by the polynomial.

[0030] In the time series clustering process, the extraction process, and the pattern calculation process, standardized data for the multiple types of objects to be provided may be used as the performance data, and in the second prediction process, data before the standardization is performed may be used as the performance data.

[0031] The item to be provided may be any one of parts constituting one or more products, a predetermined amount of materials used in manufacturing one or more products, a product belonging to one or more product groups, and a service belonging to one or more service groups.

[0032] A program according to a sixth aspect of the present disclosure causes a computer to input, into a regression learning model, a difference between actual data including a time-series actual demand count for each of a plurality of types of provision targets, the provision targets being goods or services, and first prediction data, which is data indicating a representative demand prediction value for a cluster to which the provision targets belong, obtained for each cluster classified by executing a time-series clustering process on the actual data for each of the provision targets, execute regression learning of the regression learning model, and update the regression learning model so that, during operation, the regression learning model becomes a model that inputs the difference and outputs a prediction value of the difference for a predetermined period, and is a program for executing a learning process in which, for each cluster, an extraction process is executed to extract trend component data of the cluster from the performance data, for each cluster, a pattern calculation process is executed to calculate pattern data indicating a demand pattern for the predetermined period by performing statistical processing for first data, which is data remaining after extracting the trend component data from the performance data, for the cluster, for the predetermined period, and a first prediction process is executed to calculate a sum of the pattern data and the trend component data for each cluster, thereby obtaining data indicating a representative demand prediction value of the object of service belonging to the cluster for each cluster. With such a configuration, the program makes it possible to generate a learning model used in a demand prediction system capable of predicting with sufficient accuracy the demand for the object of service, which is a product or a service, taking into account fluctuations on different time series scales for multiple types of object of service.

[0033] The specified period may be a period whose start and end dates are expressed in terms of days or months in a year, and the statistical processing may be a process of calculating statistical values ​​of the first data for the specified period in the first data for a specified number of years, for each cluster.

[0034] In the time series clustering process, the extraction process, and the pattern calculation process, standardized data for the multiple types of objects to be provided may be used as the performance data, and the difference input to the regression learning model may be a difference calculated using data before the standardization as the performance data.

[0035] The item to be provided may be any one of parts constituting one or more products, a predetermined amount of materials used in manufacturing one or more products, a product belonging to one or more product groups, and a service belonging to one or more service groups. Effect of the Invention

[0036] The present disclosure makes it possible to provide a demand forecasting system, a demand forecasting method, and a program that are capable of predicting with sufficient accuracy the demand for a provision of goods or services, taking into account fluctuations on different time series scales for multiple types of provision, as well as to provide a learning system, a learning method, and a program that are capable of generating a learning model to be used for such predictions. [Brief description of the drawings]

[0037] [Figure 1] 1 is a block diagram showing a configuration example of a demand forecasting system according to an embodiment of the present invention. [Diagram 2] FIG. 2 is a flow diagram illustrating an example of a global prediction process in the demand prediction process executed in the demand prediction system of FIG. [Diagram 3] 1. FIG. 4 is a flow diagram illustrating an example of a part-by-part prediction process in the demand prediction process executed in the demand prediction system of FIG. [Figure 4] FIG. 4 is a diagram illustrating an example of variables used in the prediction process for each part in FIG. 3. [Diagram 5] FIG. 3 is a diagram showing an example of a result of a time series clustering process in the global prediction process of FIG. 2. [Figure 6]FIG. 3 is a diagram showing an example of a result of a time series clustering process in the global prediction process of FIG. 2. [Figure 7] FIG. 3 is a diagram showing an example of a result of a time series clustering process in the global prediction process of FIG. 2. [Figure 8] FIG. 3 is a diagram showing an example of an annual demand pattern calculated in the global forecasting process of FIG. 2. [Figure 9] FIG. 3 is a diagram showing an example of an annual demand pattern calculated in the global forecasting process of FIG. 2. [Figure 10] FIG. 4 is a diagram showing an example of a prediction result for a certain part calculated in the prediction process for each part in FIG. 3. [Figure 11] FIG. 4 is a diagram showing an example of a prediction result of another certain part calculated in the prediction process for each part in FIG. 3. [Figure 12] A block diagram showing an example configuration of a learning system that generates a learning model used in the demand forecasting system of this embodiment. [Figure 13] 13 is a flow diagram illustrating an example of a learning process in the learning system of FIG. 12. [Figure 14] FIG. 2 illustrates an example of a hardware configuration of the apparatus. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0038] The present invention will be described below through embodiments of the invention, but the invention according to the claims is not limited to the following embodiments. Furthermore, not all of the configurations described in the embodiments are necessarily essential as means for solving the problems.

[0039] (Embodiment) An example of the configuration of a demand forecasting system according to the present embodiment will be described with reference to Figures 1 to 11. First, an example of the configuration of a demand forecasting system according to the present embodiment will be described with reference to Figure 1. Figure 1 is a block diagram showing an example of the configuration of a demand forecasting system according to the present embodiment (hereinafter, referred to as the present system).

[0040] The system 10 shown in FIG. 1 predicts demand for a product or service that is to be provided. In the following, the product to be provided will be described as parts such as engine parts, seat parts, driving operation part parts, and control device parts that constitute one or more products such as one or more types of automobiles. Examples of parts include at least one of parts required for a new product to be sold and parts (service parts) that are replaced during repairs or regular maintenance. Other examples of products that may be provided will be described later as alternative examples.

[0041] As shown in Figure 1, the present system 10 may include an actual data standardization unit 11, a time series clustering unit 12, a demand pattern calculation unit 13, a global prediction model generation unit 14, a residual calculation unit 15, a residual prediction model generation unit 16, and a demand prediction value calculation unit 17.

[0042] The system 10 may include each of the units 11 to 17 as a control unit. This control unit may be realized, for example, by an integrated circuit, and may be realized, for example, by a processor such as a CPU (Central Processing Unit), a working memory, and a non-volatile storage device. A control program executed by the processor is stored in this storage device, and the processor reads the program into the working memory and executes it, thereby achieving the functions of each of the units 11 to 17.

[0043] The performance data standardization unit 11 inputs performance data including the time-series actual demand numbers for each of a plurality of types of parts, standardizes the input performance data for each part, and passes it to the time-series clustering unit 12. The input performance data only needs to be time-series data representing the number of sales or reservations, etc. of the parts associated with information such as identification information indicating the parts. In addition, in order to execute processing for the plurality of types of parts, the information such as identification information indicating the parts should preferably include identification information indicating other types of parts to be taken into consideration, or a group name to be taken into consideration, etc.

[0044] The time series clustering unit 12 executes a time series clustering process on the performance data for each part received from the performance data standardization unit 11, i.e., the standardized performance data for each part, and passes the result data to the demand pattern calculation unit 13. Any method may be used for this time series clustering process.

[0045] The demand pattern calculation unit 13 executes an extraction process to extract trend component data of the cluster from the standardized performance data for each cluster classified by the time series clustering process performed by the time series clustering unit 12, and passes the extracted trend component data to the global forecast model generation unit 14. The extracted trend component data can include trend component data indicating a long-term trend.

[0046] In addition, the demand pattern calculation unit 13 executes a pattern calculation process for calculating pattern data by performing statistical processing for a predetermined period on the first data, which is the data remaining after extracting trend component data from the standardized actual data, for each cluster, and passes the pattern data to the global prediction model generation unit 14. The calculated pattern data is data indicating a demand pattern for the predetermined period. Any type of statistical processing may be used.

[0047] The global prediction model generation unit 14 executes a first prediction process to obtain first prediction data by calculating the sum of the pattern data and trend component data for each cluster based on the received trend component data and pattern data, and passes the first prediction data to the residual calculation unit 15 and the demand prediction value calculation unit 17.

[0048] The first forecast data obtained by the global forecast model generation unit 14 is data indicating a representative demand forecast value of the parts belonging to each cluster. The representative demand forecast value indicates a common demand forecast value for the parts in the cluster. The trend component data extracted as described above can include trend component data indicating a long-term trend, and therefore the first forecast data can include forecast data indicating a long-term forecast result.

[0049] The residual calculation unit 15 calculates, for each part, the difference between the pre-standardization performance data of the part and the first prediction data for the cluster to which the part belongs, and passes the result data to the residual prediction model generation unit 16.

[0050] The residual prediction model generating unit 16 generates a residual prediction model (residual model) by performing regression learning on the differences (residuals) for each part received from the residual calculating unit 15. However, in this system 10, the residual prediction model generating unit 16 can also be configured separately, and the demand forecast value calculating unit 17 can also refer to the residual model generated in advance to execute the second prediction process described later.

[0051] The demand forecast value calculation unit 17 inputs the residuals for each part received from the residual calculation unit 15 into this residual model, and executes a second forecast process to obtain second forecast data. The second forecast data is data indicating a forecast value of the residual for each part, and is data indicating a forecast value (forecast error value) of a residual that may occur in the future for each part. In particular, the second forecast data can include data indicating a forecast value of a residual that may occur in the short term in the future for each part.

[0052] Then, the demand forecast value calculation unit 17 obtains third forecast data by calculating the sum of the first forecast data and the second forecast data for each part based on the second forecast data and the first forecast data received from the global forecast model generation unit 14. The obtained third forecast data is data indicating the demand forecast value for each part.

[0053] Also, as in the example in which the present system 10 includes the performance data standardization unit 11, standardized data for multiple types of parts can be used as performance data in the time series clustering process, the extraction process, and the pattern calculation process. At this time, data before standardization can be used as performance data in the residual calculation process and the second prediction process. In this way, the residual input to the residual model in the second prediction process can use the difference calculated using data before standardization as performance data.

[0054] However, standardization of the performance data does not have to be performed inside this system 10. In that case, if an input unit for inputting the standardized performance data and the pre-standardized performance data is equipped in the demand forecasting system, the processing as exemplified above becomes possible. Of course, there may be cases where standardization of the performance data for each of multiple target types of parts is not required, in which case non-standardized performance data can be used in all processing.

[0055] Specific examples of processing in each unit 11 to 17 will be described below with reference to Figures 2 to 11. In this embodiment, by adopting the configuration described above, it becomes possible to predict demand for parts with sufficient accuracy, taking into account fluctuations on different time series scales for multiple types of parts.

[0056] A specific example of processing in the system 10 will be described. The following description will be given on the assumption that the parts are only service parts. However, even if the parts are new sales parts, both service parts and new sales parts are included, and the basic processing is the same. In addition, an example of processing in units of years will be given below. That is, in the following processing example, the predetermined period is a period whose start date and end date are expressed by a day or month in a year, and the statistical processing is an example of processing in which the statistical value of the first data is calculated for each cluster for a predetermined period in the first data for a predetermined number of years. However, the predetermined period and the statistical processing are not limited to this example, and an appropriate predetermined period and statistical processing may be selected depending on the object of provision. In addition, when the predetermined period is a period expressed by months, for example, the first day of the starting month can be used as the start date and the first or last day of the ending month can be used as the end date, or the middle day of the starting month (for example, the 16th) can be used as the start date and the middle day of the ending month (for example, the 15th) can be used as the end date. However, it is not limited to these days, and it is only necessary to determine in advance which days of the specified month to adopt as the start date and the end date.

[0057] As can be seen from the explanation of FIG. 1, the processing in the present system 10 can be a two-stage hierarchical processing consisting of a global prediction processing and a prediction processing for a single part.

[0058] In global prediction processing, for example, the similarity of time series is calculated, and the time series data is classified into several clusters with similar features such as periodicity due to seasonal fluctuations, and a cluster center demand pattern is calculated as information representative of each cluster, and a demand prediction that captures global characteristics is performed from this demand pattern.

[0059] In the prediction process for a single part, the characteristics of the single part are taken into consideration. That is, in the prediction process for a single part, for example, the residual between the demand forecast based on the characteristics obtained by the time series clustering process and the actual results is modeled. Then, in the prediction process for a single part, the demand forecast for the single part is performed by taking the sum of the global forecast result obtained in the global forecast process and the forecast result of the residual.

[0060] First, an example of a global prediction process will be described in detail with reference to Fig. 2. Fig. 2 is a flow diagram for explaining an example of a global prediction process in a demand prediction process executed in the present system 10.

[0061] In addition to seasonal fluctuations over a year (12 months) and shorter-term fluctuations, parts also have a noise component called random walk. Therefore, as a global forecasting process, similar parts are aggregated and the noise is removed by taking the average, and the aggregated demand trends are used for forecasting.

[0062] First, the time series clustering unit 12 executes a time series clustering process for time series clustering the standardized performance data (step S1). In this time series clustering process, parts with similar time series changes are classified into the same group. Here, an example is given of a time series clustering process using the Ward method. However, the clustering method for time series data is not limited to the Ward method, and for example, a shortest distance method or a group average method using differences in distance functions for measuring similarity can also be applied. In addition, as a clustering method for time series data, a non-hierarchical method such as the Kmeans method can also be applied. In this embodiment, an example is given of using the Ward method, which does not require the number of clusters to be set in advance and has high classification sensitivity.

[0063] Ward's method is analyzed in the following steps. 1.Let the number of data be N, and create N clusters each containing only one piece of data. 2. Let the clusters be C1 and C2, and define their distance function similarity as d(C1, C2). 3. Merge the two clusters with the smallest d(C1,C2). 4. Analyze sequentially until all objects are divided into two clusters, and create a hierarchical dendrogram.

[0064] Here, in the Ward method, the distance function d(C1, C2) is defined as the following formula (1). However, in formula (1), E(C1) is defined by formula (2). Here, c1 is the center of gravity of C1, the overline of y represents the standardized time series, and is calculated using the Euclidean distance D, which is used as the distance between vectors based on c1.

[0065]

number

[0066] Next, the time series clustering unit 12 extracts trend components and performs polynomial approximation to determine whether or not a trend component is included (step S2). If a trend component is included, that is, if the answer is YES in step S2, the time series clustering unit 12 extracts trend component T (step S3).

[0067] If it is determined in step S2 that a trend component is included, the time series clustering unit 12 checks the determination result as necessary after the processing of step S3 (step S4), and if a trend component is included, that is, if the result of step S4 is YES, the process returns to step S1 and performs the time series clustering process again. In this way, in this system 10, in order to extract characteristics of periodic fluctuations such as seasonality as a demand pattern, the trend component is extracted from clusters that include a trend component from the result of the time series clustering process, and the time series clustering process is performed again.

[0068] For this purpose, the system 10 performs decomposition of the time series data. In general, the time series data U can be defined as an additive expression of the trend T, the seasonal variation S, and the surplus term R. In this embodiment, the trend component T is extracted using STL decomposition (Seasonal Decomposition Of Time Series By Loess), which can decompose the time series data U into T, S, and R, respectively. Compared with other decomposition methods, the STL decomposition does not require constraints on seasonal variation, is robust against outliers, and allows robust decomposition. Therefore, it is useful for time series data of spare parts that also include unexpected order records. However, in this embodiment, the decomposition of the time series data can also be performed using a method other than the STL decomposition, and in particular, if the predetermined period is different from the example described here, a corresponding decomposition can be performed.

[0069] If the result in step S4 is NO, the time series clustering unit 12 performs polynomial approximation on the extracted trend component T to calculate an approximation formula (step S5). The reason for performing such calculation is that extrapolation is necessary to predict future demand. Here, as an example, a third-order polynomial regression is performed to calculate an approximation function, but the present invention is not limited to this.

[0070] In this way, the system 10 can execute a determination process for determining whether or not a trend component is included for each cluster on the post-extraction data, which is data after the data of the trend component is extracted in the extraction process. Then, in the system 10, if it is determined as a result of the determination process that at least one cluster contains a trend component, the system 10 executes the time series clustering process and the extraction process on the post-extraction data instead of the performance data, until it is determined in the determination process that the data does not contain a trend component.

[0071] If step S2 is NO, that is, if the extraction of trend components is completed, the demand pattern calculation unit 13 generates a demand pattern. In the demand pattern generation process, for example, a time series representative of each cluster is found and utilized from the result of the obtained time series clustering process. As a demand pattern calculation method, various methods can be adopted, such as a method of calculating a time series that minimizes the sum of dissimilarities between the distribution in the cluster and the time series in other clusters, or a method of using the average value in the cluster as a representative time series.

[0072] In this embodiment, as an example, the demand pattern calculation unit 13 calculates the average value of the order records of parts belonging to each cluster as a representative time series (step S6), takes the average for each month from this representative time series, and calculates the annual demand pattern (overline of Z) using the following equation (3) (step S7).

[0073]

number

[0074] Here, k is the cluster number, l is the year, m is the month, L is the number of years of the learning period, and N k is the number of part numbers in the cluster. i,k,l,m The overlines represent the real values ​​of the standardized time series data that belong to the cluster.

[0075] In this manner, the demand pattern calculation unit 13, as a pattern calculation process, extracts trend component data from the post-extraction data in place of actual data when it is determined in the determination process that the data does not contain a trend component, sets the remaining data as first data, and performs statistical processing on the first data for a predetermined period, thereby calculating pattern data.

[0076] After the processing of steps S5 and S7 is completed, the global forecast model generation unit 14 finally performs a global demand forecast for each part during the analysis period based on the sum of the annual demand pattern (overline of Z) and the approximation function of the trend component (hat (^) of T) (step S8), and the global forecast processing is terminated.

[0077] In this way, the first prediction process, which is a global prediction process, includes an approximation process that approximates the trend component data extracted in the extraction process to a polynomial, and by calculating the sum of the pattern data and the data indicated by the polynomial for each cluster, the first prediction data can be obtained for each cluster.

[0078] The global forecasting process described above can represent the long-term linear dependency and periodic seasonal fluctuations of service parts, and can make forecasts that take similar parts into account. However, this alone makes it difficult to predict future demand for a single part with sufficient accuracy. This is because the demand characteristics differ for each part, with seasonal fluctuations of one or two months differing, and even smaller fluctuations being mixed in.

[0079] Therefore, the present system 10 performs the above-mentioned second prediction process and a process of calculating the sum of the first prediction data and the second prediction data as a prediction process for a single part. Specifically, the present system 10 models the difference between the global prediction based on the polynomial regression of the trend and the sum of the demand pattern on an annual basis and each actual result of each part as a prediction process for a single part.

[0080] An example of a forecasting process for a single part will be described in detail below with reference to Fig. 3 to Fig. 11. Fig. 3 is a flow diagram for explaining an example of a forecasting process for each part in the demand forecasting process executed by the present system 10.

[0081] First, the residual calculation unit 15 calculates the difference (residual) between the global prediction and the actual result (step S11). i,m can be defined, for example, as the following equation (4): i,m is the order record, and Z k The overline of is the annual demand pattern obtained by clustering. i The hat (^) in the above table indicates the trend components predicted by polynomial regression, and each is reverted to its unstandardized value.

[0082]

number

[0083] Next, the residual prediction model generation unit 16 inputs the residuals for each part received from the residual calculation unit 15 to a residual prediction model (residual model). This residual model is a learning model for feature extraction processing to extract feature vectors from the residuals and for regression learning of the extracted feature vectors. Then, the residual prediction model generation unit 16 extracts feature vectors of the residuals from the input residuals (step S12), and updates the residual model by regression learning of the feature vectors of the residuals (step S13).

[0084] This residual model can be defined as the above formula (5). Here, f is a function for predicting the residual. The function f can be, but is not limited to, a neural network or a gradient boosting tree.

[0085] Also, x n defines time terms such as past residuals and months. Also, x1, x2, , x n can have design variables such as those shown in Fig. 4. Fig. 4 is a diagram showing an example of variables used in the prediction process for each part in Fig. 3. In Fig. 4, past residual results from 4 months ago to 13 months ago are set as design variables, and the number of months and values ​​obtained by making the number of months into cyclic variables using trigonometric functions are set in order to make the relationship between the results and time a feature amount (feature vector).

[0086] Next, the demand forecast value calculation unit 17 inputs the residual of the target part into the residual model to obtain a forecast value of the residual (step S14). Next, the demand forecast value calculation unit 17 calculates and outputs the demand forecast value of the part by calculating the sum of the forecast value of the residual, which is the predicted residual, and the globally predicted value (step S15), and the process ends. The forecast value of equation (5) is expressed as e i,m If we set the hat (^) of the demand forecast, the final demand forecast value to be output can be expressed by the above equation (6).

[0087] Next, the results of demand forecasting for certain service parts by the system 10 will be described with reference to Fig. 5 to Fig. 11. As an application example, time-series clustering processing is performed on approximately 9,000 service parts with seasonal fluctuations, and demand forecasting is performed on 910 service parts, which is approximately 10% of the total.

[0088] First, an example of a demand forecast result will be given with reference to Figs. 5 to 9. Figs. 5 to 7 are diagrams showing an example of the result of the time series clustering process in the global forecast process of Fig. 2, and Figs. 8 and 9 are diagrams showing an example of an annual demand pattern calculated in the global forecast process of Fig. 2. In Figs. 5 to 9, the horizontal axis shows the time (scale is in 6-month units), and the vertical axis shows the order amount (standardized order amount) of each part in the cluster. Note that, for simplification, Figs. 5 to 9 omit much more than 910 pieces of data to be shown, and the data is deformed to a degree that allows the data to be qualitatively understood.

[0089] First, by performing time series clustering processing on the standardized time series performance data, a total of three clusters are generated as shown in each of Figures 5 to 7. Then, as a result of extracting trend components from the cluster shown in Figure 5 and calculating the demand pattern on an annual basis, a demand pattern as shown by the solid line in Figure 8 is generated. Comparing Figures 5 and 8, it can be seen that no trend component was present in this cluster. On the other hand, as a result of extracting trend components from the cluster shown in Figure 6 and calculating the demand pattern on an annual basis, a demand pattern as shown by the solid line in Figure 9 is generated. In Figure 6, there was an overall downward trend (trend component), but in Figure 9, it can be seen that this trend component has already been extracted.

[0090] Next, the residuals were calculated, a residual model was constructed, and accuracy verification was performed. The results will be described with reference to Figs. 10 and 11. Fig. 10 is a diagram showing an example of a prediction result for a certain part calculated in the prediction process for each part in Fig. 3, and Fig. 11 is a diagram showing an example of a prediction result for another certain part calculated in the prediction process for each part in Fig. 3. In Figs. 10 and 11, the horizontal axis shows time (in six-month units), and the vertical axis shows the order volume, with the solid line showing the actual order volume and the dotted line showing the predicted order volume.

[0091] As design variables, the past residual results from 4 months ago to 13 months ago are set as shown in Figure 4. Furthermore, as design variables, in order to treat the relationship between results and time as a feature quantity, the number of months and the value obtained by making the number of months a repeating variable using a trigonometric function are set.

[0092] In this example, in order to evaluate the generalization performance of the residual model, the prediction period was divided into three parts as shown in CV1, CV2, and CV3 below, and time-series cross-validation was performed. In CV1, the learning period was from August 2017 to May 2020, and the testing period was from June 2020 to September 2020. In CV2, the learning period was from August 2017 to September 2020, and the testing period was from October 2020 to January 2021. In CV3, the learning period was from August 2017 to January 2021, and the testing period was from February 2021 to May 2021.

[0093] In addition, the following formula (7) was used as the evaluation index: By using formula (7), it is possible to calculate the forecast error standardized by the average of the actual results for the past year from the forecast month.

[0094]

number

[0095] Next, the allowable error was set to -0.3≦Error≦0.5, and an evaluation was conducted based on the percentage of parts for which the prediction accuracy was within the allowable error range for the entire test period out of the 910 parts targeted for prediction. The allowable error was set to the above range because it is equivalent to the error after manual correction, reduces the risk of stockouts, and allows for management of excess inventory up to the appropriate inventory level +50.

[0096] For comparison, a regression model was constructed for each individual part, and the predicted results were used. As a result of the evaluation, it became clear that while the accuracy rate of the comparison method was 0%, the accuracy of the demand forecast by this system 10 was improved, and it was possible to predict within the allowable error range for 26.7% of the total parts. Figures 10 and 11 show the prediction results for two parts for which demand forecasting was performed by this system 10 and predictions were made within the allowable error range, and it can be seen that demand forecasting was performed with high prediction accuracy in both cases.

[0097] As described above, according to this embodiment, it is possible to predict the demand for parts with sufficient accuracy by taking into account fluctuations on different time series scales for multiple types of parts. More specifically, according to this embodiment, not only can the demand for a single part be predicted by taking into account long-term trends in the demand for multiple types of parts and periodic fluctuations due to seasons, etc., but also future demand can be predicted with sufficient accuracy by taking into account the demand characteristics of a single part.

[0098] The effects of this embodiment will be considered with respect to an example of a specific application scene of this embodiment. For example, in order to gain the trust of customers, service parts need to be supplied in a timely manner in response to parts replacement due to accidents and regular replacement of consumables, not to mention replacement of parts due to manufacturing responsibility. Service parts have not only trend components due to aging but also periodic fluctuations such as seasonal fluctuations. Furthermore, depending on the characteristics of each part, they have periodicities and noise components that are smaller than seasonality. Therefore, for appropriate inventory management, it is necessary to predict the monthly demand number with high accuracy. However, due to such problems, there is a possibility that it may be difficult to accurately predict the demand for individual parts, including short-term time series fluctuations, and that accurate clustering cannot be performed for demand characteristics with different scales.

[0099] In contrast, in this embodiment, in order to utilize information on similar characteristics between parts, a prediction method is used that combines global prediction, which extracts demand characteristics by clustering standardized time series data, and prediction by modeling residuals for individual parts in each cluster. That is, in this embodiment, trend components are extracted, and a demand pattern for each part with high similarity in demand characteristics is calculated, global demand prediction is performed, and further, prediction of each part is performed by predicting characteristics specific to each service part from the residual between the result of the global demand prediction and the actual performance. Therefore, in this embodiment, it is possible to reduce the possibility that accurate prediction becomes difficult or accurate clustering cannot be performed due to the above problems, and the risk of shortages and excess inventory of service parts can be reduced.

[0100] Next, an example of a learning system that generates the above-mentioned learning model (residual model) will be described with reference to Fig. 12 and Fig. 13. Fig. 12 is a block diagram showing an example of a configuration of a learning system that generates a learning model used in the demand forecasting system 10. Fig. 13 is a flow diagram for explaining an example of a learning process in the learning system of Fig. 12.

[0101] As shown in Fig. 12, the learning system 50 can include a control unit 51, a storage unit 52, and an input unit 53. The learning system 50 corresponds to the system described as the residual prediction model generation unit 16 in Fig. 1, and can be generated using a computer such as a PC for AI (Artificial Intelligence) learning. However, the learning system 50 may be configured as a single device or may be configured by distributing functions among multiple devices.

[0102] The control unit 51 controls the entire learning system 50. The control unit 51 can be realized, for example, by an integrated circuit, and can be realized, for example, by a processor such as a CPU, a working memory, and a non-volatile storage device. The control program executed by the processor is stored in this storage device, and the processor reads the program into the working memory and executes it, thereby fulfilling the function of the control unit 51. The control program can include a learning program that executes learning. Note that the storage device can also utilize the memory unit 52.

[0103] The memory unit 52 is a storage device that stores residual data 52a, which is data indicating residuals, and a learning model 52b, which can be stored in, for example, a database format. The input unit 53 inputs a data set required for learning, at least the residual data 52a, and stores it in the memory unit 52 so that it can be referenced during learning. When learning includes supervised learning, the residual data 52a should also include correct answer data.

[0104] 13, the processing by the learning system 50 is performed by first inputting the residual data 52a from the input unit 53 to the control unit 51 (step S21) and executing regression learning (step S22). For example, the control unit 51 can read the residual data 52a during learning, execute learning based on the learning model 52b, and update the learning model 52b.

[0105] The learning model 52b and the regression learning using it have been explained with reference to Figure 3 etc., and detailed explanations will be omitted as long as a learning model to be used in the second prediction process can be generated, but the control unit 51 will execute the following processing.

[0106] The control unit 51 inputs the residuals between the result data including the time-series actual demand count for each of the multiple types of parts and the first prediction data, which is data indicating a representative demand prediction value for the cluster to which the part belongs (a representative demand prediction value for the part) obtained for each cluster classified by executing a time-series clustering process on the result data for each part, to the regression learning model. Then, the control unit 51 executes regression learning of this regression learning model and updates the regression learning model. This update is executed so that the regression learning model becomes a model that inputs the residuals and outputs a prediction value of the residual for a predetermined period during operation.

[0107] Here, the first forecast data is obtained as data indicating representative demand forecast values ​​of parts belonging to each cluster by executing the extraction process, the pattern calculation process, and the first forecast process as described above. The extraction process is a process of extracting trend component data of each cluster from the performance data. The pattern calculation process is a process of calculating pattern data indicating a demand pattern for a predetermined period by performing statistical processing for a predetermined period on the first data, which is the data remaining after extracting trend component data from the performance data for each cluster. The first forecast process is a process of calculating the sum of the pattern data and the trend component data for each cluster.

[0108] With this configuration, the learning system 50 is able to generate a learning model for use in a demand forecasting system that can predict demand for parts with sufficient accuracy, taking into account fluctuations on different time series scales for multiple types of parts.

[0109] (Alternative examples, etc.)

[0110] In the above embodiment, the parts constituting one or more products are exemplified as automobiles as the object of provision, but the product is not limited to automobiles and may be any product such as other types of equipment, other types of daily necessities, other types of food, etc. For example, if the product is a snack, the ingredients and packaging of the snack may be exemplified as the parts.

[0111] In the above embodiment, the parts constituting one or more products are given as an example of the object to be provided. However, the object to be provided may be, for example, any one of (a) parts constituting one or more products, (b) a predetermined amount of material used in manufacturing one or more products, (c) a product belonging to one or more product groups, and (d) a service belonging to one or more service groups.

[0112] An example of the above (b) is a predetermined amount of material such as paint for a 18-liter can, which is used when manufacturing one or more products such as one or more types of automobiles. An example of the above (c) is a product belonging to one or more product groups, such as chocolate confectionery, and in this case, examples of the product include almond chocolate and chocolate bars, and can be specifically distinguished by product name. An example of the above (d) is a service belonging to one or more service groups, such as a service group for transporting people, and in this case, examples of the service include trains, express trains, buses, and taxis. This makes it possible to predict the demand for services using MaaS (Mobility as a Service) vehicles, for example. Of course, various examples other than the above-mentioned examples can be applied to the above (a) to (d).

[0113] Furthermore, not only the learning system described above, but also the demand reservation system described above can be constructed as a system in which functions are distributed among multiple devices, or as a single device. Either of these multiple devices or a single device can have, for example, the following hardware configuration. Figure 14 is a diagram showing an example of the hardware configuration of such a device.

[0114] 14 may each include a processor 101, a memory 102, and an interface 103. For example, when the device 100 is a computer, the interface 103 may include an interface with an operation device that accepts user operations, an interface with a display device that displays information, a communication interface for sending and receiving information, and the like.

[0115] The processor 101 may be, for example, a microprocessor, a microprocessor unit (MPU), or a CPU. The processor 101 may include multiple processors. The memory 102 is configured, for example, by a combination of a volatile memory and a non-volatile memory. The functions of the above-mentioned multiple devices or a single device are realized by the processor 101 reading a control program stored in the memory 102 and executing the program while exchanging necessary information via the interface 103.

[0116] The program includes instructions (or software code) that, when loaded into a computer, cause the computer to perform one or more functions described in the embodiments. The program may be stored in a non-transitory computer-readable medium or a tangible storage medium. By way of example and not limitation, the computer-readable medium or tangible storage medium includes random-access memory (RAM), read-only memory (ROM), flash memory, solid-state drive (SSD) or other memory technology, CD-ROM, digital versatile disc (DVD), Blu-ray® disk or other optical disk storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage device. The program may be transmitted on a transitory computer-readable medium or communication medium. By way of example and not limitation, the transitory computer-readable medium or communication medium includes electrical, optical, acoustic, or other form of propagated signal.

[0117] The present invention is not limited to the above-described embodiment, and can be modified as appropriate without departing from the spirit and scope of the present invention. [Explanation of symbols]

[0118] 10 Demand forecasting system 11 Performance Data Standardization Department 12 Time Series Clustering Section 13 Demand pattern calculation section 14 Global prediction model generation unit 15 Residual calculation section 16 Residual prediction model generation unit 17 Demand forecast value calculation section 50 Learning System 51 Control section 52 Storage section 52a Residual data 52b Learning Model 53 Input section 100 devices 101 Processor 102 Memory 103 Interface

Claims

1. A demand forecasting system including a computer, the computer comprising: inputting actual data including a time-series actual demand number for each of a plurality of types of provision targets, which are goods or services, and executing a time-series clustering process on the input actual data for each of the provision targets; executing an extraction process for extracting trend component data of the cluster from the performance data for each cluster classified by the time series clustering process; executing, for each cluster, a pattern calculation process for calculating pattern data indicating a demand pattern for a predetermined period by performing a statistical process on first data, which is data remaining after extracting data of the trend component from the performance data, for the predetermined period; executing a first prediction process for obtaining, for each cluster, first prediction data indicating a representative demand prediction value of the object to be provided that belongs to the cluster by calculating a sum of the pattern data and the trend component data for each cluster; A second prediction process is executed to obtain second prediction data, which is data indicating a predicted value of the difference for each of the objects to be provided, by inputting a difference between the performance data of the object to be provided and the first prediction data for the cluster to which the object to be provided belongs to a learning model that has undergone regression learning of the difference for the object to be provided, and calculating a sum of the first forecast data and the second forecast data for each of the provision targets to obtain third forecast data which is data indicating a demand forecast value for each of the provision targets; Demand forecasting system.

2. The computer, performing a determination process for determining whether or not a trend component is included in each of the clusters with respect to post-extraction data, which is data after the data of the trend component is extracted in the extraction process; When it is determined that at least one of the clusters contains a trend component as a result of the determination process, the time series clustering process and the extraction process are executed on the extracted data instead of the performance data until it is determined that the extracted data does not contain a trend component in the determination process; the pattern calculation process, when it is determined in the determination process that the data does not contain a trend component, extracts data of the trend component from the extracted data instead of the performance data, sets the remaining data as the first data, and calculates the pattern data by performing a statistical process on the first data for the predetermined period; The demand forecasting system according to claim 1 .

3. The predetermined period is a period whose start and end dates are expressed in days or months of a year, The statistical processing is a process of calculating a statistical value of the first data for the predetermined period of the first data for a predetermined number of years for each cluster. The demand forecasting system according to claim 1 or 2.

4. the first prediction process includes an approximation process of approximating the data of the trend component extracted in the extraction process to a polynomial, and a sum of the pattern data and the data represented by the polynomial is calculated for each of the clusters to obtain the first prediction data for each of the clusters. The demand forecasting system according to any one of claims 1 to 3.

5. In the time series clustering process, the extraction process, and the pattern calculation process, standardized data for the plurality of types of objects to be provided is used as the performance data, and in the second prediction process, data before the standardization is used as the performance data. The demand forecasting system according to any one of claims 1 to 4.

6. The subject of provision is: Parts that make up one or more products; A quantity of material to be used in producing one or more products; Products belonging to one or more product groups; and Services belonging to one or more service groups; Any one of the following: The demand forecasting system according to any one of claims 1 to 5.

7. A computer comprising: inputting actual data including a time-series actual demand number for each of a plurality of types of provision targets, which are goods or services, and executing a time-series clustering process on the input actual data for each of the provision targets; executing an extraction process for extracting trend component data of the cluster from the performance data for each cluster classified by the time series clustering process; executing, for each cluster, a pattern calculation process for calculating pattern data indicating a demand pattern for a predetermined period by performing a statistical process on first data, which is data remaining after extracting data of the trend component from the performance data, for the predetermined period; executing a first prediction process for obtaining, for each cluster, first prediction data indicating a representative demand prediction value of the object to be provided that belongs to the cluster by calculating a sum of the pattern data and the trend component data for each cluster; A second prediction process is executed to obtain second prediction data, which is data indicating a predicted value of the difference for each of the objects to be provided, by inputting a difference between the performance data of the object to be provided and the first prediction data for the cluster to which the object to be provided belongs to a learning model that has undergone regression learning of the difference for the object to be provided, and calculating a sum of the first forecast data and the second forecast data for each of the provision targets to obtain third forecast data which is data indicating a demand forecast value for each of the provision targets; Demand forecasting methods.

8. On the computer, inputting actual data including a time-series actual demand number for each of a plurality of types of provision targets, which are goods or services, and executing a time-series clustering process on the input actual data for each of the provision targets; executing an extraction process for extracting trend component data of the cluster from the performance data for each cluster classified by the time series clustering process; executing, for each cluster, a pattern calculation process for calculating pattern data indicating a demand pattern for a predetermined period by performing a statistical process on first data, which is data remaining after extracting data of the trend component from the performance data, for the predetermined period; executing a first prediction process for obtaining, for each cluster, first prediction data indicating a representative demand prediction value of the object to be provided that belongs to the cluster by calculating a sum of the pattern data and the trend component data for each cluster; A second prediction process is executed to obtain second prediction data, which is data indicating a predicted value of the difference for each of the objects to be provided, by inputting a difference between the performance data of the object to be provided and the first prediction data for the cluster to which the object to be provided belongs to a learning model that has undergone regression learning of the difference for the object to be provided, and calculating a sum of the first forecast data and the second forecast data for each of the provision targets to obtain third forecast data which is data indicating a demand forecast value for each of the provision targets; A program for executing demand forecast processing.

Images