Production planning system, learning device, and inference device

The production planning system addresses the challenge of managing small quantity production by using data-driven prediction models to classify products and select optimal forecasting methods, resulting in improved accuracy and reduced production adjustments.

WO2024236615A1PCT designated stage expired Publication Date: 2024-11-21MITSUBISHI ELECTRIC CORP
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
PCT/JP2023/017841
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-12
Publication Date
2024-11-21

AI Technical Summary

Technical Problem

Conventional production planning systems face challenges in creating appropriate production plans for a large number of products in small quantities, as they struggle to accurately forecast demand and manage inventory effectively.

Method used

A production planning system that acquires past sales performance data to predict demand for each product model, classifies products into subcategories, and selects the optimal prediction model with the minimum prediction error rate, using a combination of overall, subcategory, and model-specific prediction units, along with machine learning, to formulate a production plan.

Benefits of technology

The system enables accurate demand forecasting and stable production planning by selecting the most accurate prediction model, reducing production adjustments and improving prediction accuracy in response to economic fluctuations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure JP2023017841_21112024_PF_FP_ABST
    Figure JP2023017841_21112024_PF_FP_ABST
Patent Text Reader

Abstract

A production planning system (1) comprises: a performance data acquisition unit (11) that acquires past sales performance data of each of a plurality of products; an individual model prediction unit (24) that predicts the demand for each model of the products on the basis of the sales performance data acquired by the performance data acquisition unit (11); a small classification group prediction unit (23) that, on the basis of the sales performance data acquired by the performance data acquisition unit (11), classifies the products into small classification groups including a plurality of models and that predicts the demand for each small classification group; an entirety prediction unit (21) that predicts the demand for the entirety of the products on the basis of the sales performance data acquired by the performance data acquisition unit (11); a selection unit (13) that selects an optimal prediction model which has the lowest prediction error rate for the past one year among demand predictions predicted by a prediction unit (12) having the individual model prediction unit (24), the small classification group prediction unit (23), and the entirety prediction unit (21); and a production planning unit (14) that designs a production plan on the basis of the optimal prediction model selected by the selection unit (13).
Need to check novelty before this filing date? Find Prior Art

Description

Production planning system, learning device and inference device

[0001] The present disclosure relates to a production planning system, a learning device, and an inference device that create production plans for small-lot, multi-model products.

[0002] A conventional production planning system is known that manages inventory by calculating the number of days of inventory based on the relationship between the inventory quantity of managed items, actual sales volume, and product-specific correction coefficients (see, for example, Patent Document 1).

[0003] Japanese Patent Application Laid-Open No. 2001-229319

[0004] However, conventional production planning systems have a problem in that when producing a large number of different models in small quantities, it is unclear how to create an appropriate production plan for each model.

[0005] The present disclosure has been made in view of the above, and aims to provide a production planning system that can select an appropriate demand forecasting model.

[0006] In order to solve the above-mentioned problems and achieve the object, a production planning system according to the present disclosure is a production planning system that plans the production volume of a plurality of products in a predetermined unit period, and includes: a performance data acquisition unit that acquires past sales performance data for each of the plurality of products; a model-by-model prediction unit that predicts demand for each product model based on the sales performance data acquired by the performance data acquisition unit; a sub-category prediction unit that classifies the products into sub-categories including a plurality of models based on the sales performance data acquired by the performance data acquisition unit and predicts demand for each sub-category; and an overall prediction unit that predicts demand for all products based on the sales performance data acquired by the performance data acquisition unit. The production planning system according to the present disclosure further includes: a selection unit that selects an optimal prediction model that has the smallest prediction error rate for the past year from among the demand forecasts predicted by the prediction unit having the model-by-model prediction unit, the sub-category prediction unit, and the overall prediction unit; and a production planning unit that formulates a production plan based on the optimal prediction model selected by the selection unit.

[0007] The production planning system according to the present disclosure has the effect of being able to select an appropriate demand forecasting model.

[0008] FIG. 1 is a diagram showing the configuration of a production planning system according to the first embodiment;

[0009] A production planning system, a learning device, and an inference device according to an embodiment will be described in detail below with reference to the drawings.

[0010] 1 is a diagram showing the configuration of a production planning system 1 according to embodiment 1. The production planning system 1 is a system that plans the production volume of a plurality of products in a predetermined unit period, and includes a performance data acquisition unit 11, a prediction unit 12, a selection unit 13, and a production planning unit 14. The prediction unit 12 includes an overall prediction unit 21, a major classification prediction unit 22, a minor classification prediction unit 23, a model-specific prediction unit 24, and a machine learning prediction unit 25.

[0011] The performance data acquisition unit 11 acquires past sales performance data for each of a plurality of products. The overall prediction unit 21 predicts demand for all products based on the sales performance data acquired by the performance data acquisition unit 11. The major category prediction unit 22 classifies products into major categories including multiple minor categories based on the sales performance data acquired by the performance data acquisition unit 11, and predicts demand for each major category. The minor category prediction unit 23 classifies products into minor categories including multiple models based on the sales performance data acquired by the performance data acquisition unit 11, and predicts demand for each minor category. The model-specific prediction unit 24 predicts demand for each product model based on the sales performance data acquired by the performance data acquisition unit 11. The machine learning prediction unit 25 predicts demand using machine learning. The prediction unit 12 predicts demand based on an annual prediction value estimated based on changes in annual sales performance values ​​over time and a seasonal fluctuation value that is the difference between the average sales performance value and the average performance value for a unit period.

[0012] The selection unit 13 selects an optimal forecast model that has the smallest forecast error rate for the past year from among the demand forecasts made by the forecasting unit 12, which has an overall forecasting unit 21, a sub-category forecasting unit 23, and a model-specific forecasting unit 24. The production planning unit 14 formulates a production plan based on the optimal forecasting model selected by the selection unit 13.

[0013] 2 is a schematic diagram of model classification according to the first embodiment. The production planning system 1 is used to plan the number of units to be produced for each model when multiple models are produced in the same factory. For example, the production planning system 1 is used when several thousand models of products are produced in total, several hundred thousand units per month.

[0014] The multiple models are classified into minor categories and major categories. The minor categories are formed into groups containing about 5 to 20 models, preferably about 10 models. The major categories are formed into groups containing about 5 to 20 minor categories, preferably about 10 models, and about 100 models. Models are types of products with the same function, shape, specifications, etc. Major categories are functional classification groups in which models are classified by function. Minor categories are specification classification groups in which the major categories are classified by specifications.

[0015] As shown in FIG. 2 , the overall product P includes major categories P1, P2, and P3. Major category P1 includes subcategories P11, P12, P13, and P14. Subcategory P11 includes models P111, P112, P113, and P114. Subcategory P12 includes models P121, P122, and P123. Subcategory P13 includes models P131, P132, P133, and P134. Subcategory P14 includes models P141, P142, and P143. Major category P2 includes subcategory P21, P22, and P23. Subcategory P21 includes models P211, P212, P213, and P214. The minor category P22 includes model P221, model P222, and model P223. The minor category P23 includes model P231, model P232, model P233, and model P234. The major category P3 includes subcategories P31, P32, P33, and P34. The minor category P31 includes model P311, model P312, model P313, and model P314. The minor category P32 includes model P321, model P322, and model P323. The minor category P33 includes model P331, model P332, model P333, and model P334. The minor category P34 includes model P441, model P442, and model P443. In other words, each minor category includes multiple models, and each major category includes multiple minor categories.

[0016] In the case of a ventilation fan production system, the major categories include electric fans, duct fans, pressurized ventilation fans, heat exchange ventilation systems, and air conditioning fans. The minor categories include living room, office, kitchen, and bathroom categories. Each minor category may consist of, for example, about 10 models with different rated airflow rates.

[0017] Fig. 3 is a diagram showing an example of monthly production numbers of a specific model according to embodiment 1. Fig. 4 is a diagram showing trends in annual production numbers of a specific model according to embodiment 1. Fig. 4 shows trends in annual production numbers of the specific model featured in Fig. 3. Fig. 5 is a diagram showing differences from the monthly annual average values ​​of the specific model according to embodiment 1. Fig. 5 shows differences from the monthly annual average values ​​of the specific model featured in Fig. 3.

[0018] FIG. 6 is a flowchart showing the steps of the operation in the process of predicting the monthly production volume of each model according to the first embodiment. In the performance data acquisition process S1, the performance data acquisition unit 11 acquires data indicating the monthly production volume for each model over the past 10 years. In the aging estimation process S2, the prediction unit 12 calculates the annual production volume for each model over the past 10 years as shown in FIG. 4 and predicts the annual production volume for the next fiscal year based on the aging volume. For example, the prediction unit 12 predicts the annual production volume for the next fiscal year by quadratic approximation based on the annual production volume for the past 10 years. As an approximation method, for example, linear approximation or logarithmic approximation may be used. As shown in FIG. 4, a trend of change is often observed in the aging volume, and the prediction unit 12 can predict the annual production volume for the next fiscal year by quadratic approximation or the like.

[0019] In the monthly fluctuation estimation step S3, the prediction unit 12 calculates the difference between the monthly production volume and the annual average production volume for the target model, as shown in Figure 5, and then averages these monthly values ​​over 10 years to calculate the difference from the annual average value. Regarding monthly fluctuations in production volume, as shown in Figure 3, for example, for the target model shown in Figure 3, seasonal periodic fluctuations are observed, such as low production in August and high production in March. This monthly fluctuation estimation method corresponds to zero-order approximation (constant approximation) with the time axis as a variable. Since monthly fluctuations also change over time, the approximation method can be selected from constant approximation, linear approximation, exponential approximation, logarithmic approximation, and polynomial approximation.

[0020] In the monthly production volume prediction step S4, the prediction unit 12 predicts the monthly production volume based on the estimated aging change value and the estimated monthly fluctuation. In this way, the prediction unit 12 makes a prediction for each model.

[0021] Fig. 7 is a flowchart showing the procedure of the overall operation of the production plan according to embodiment 1. In a model-by-model prediction step S11, the prediction unit 12 performs the model-by-model prediction shown in Fig. 6. In the model-by-model prediction step S11, the model-by-model prediction unit 24 performs all steps in the flowchart shown in Fig. 6.

[0022] In the subcategory prediction step S12, the subcategory prediction unit 23 summarizes the number of units for each subcategory and predicts the monthly production volume in the same manner as the model-specific prediction. After predicting the production volume for each subcategory, the subcategory prediction unit 23 allocates the production volume for each subcategory to the production volume for each model based on the average production ratio for each model over a 10-year period, for example. This model-specific allocation estimation method corresponds to zero-order approximation (constant approximation) with the time axis as a variable. Since the model-specific allocation also changes over time, the approximation method can be selected from constant approximation, linear approximation, exponential approximation, logarithmic approximation, and polynomial approximation.

[0023] In the major classification prediction step S13, the major classification prediction unit 22, similar to the minor classification prediction, summarizes the number of units for each major classification, predicts the number of units to be produced for each major classification, and allocates the number of units to be produced for each model.

[0024] In the overall prediction step S14, the overall prediction unit 21 summarizes the total number of units to be produced, predicts the number of units to be produced, and then allocates the total number of units to the number of units to be produced for each model.

[0025] In the selection step S15, the selection unit 13 selects the model with the smallest prediction error rate for the past year from among the model-specific prediction, small category prediction, large category prediction, and overall prediction. The prediction unit 12 performs an error evaluation prediction for FY20 using model-specific prediction, small category prediction, large category prediction, and overall prediction based on past performance data for the 10 years from FY2010 to FY2019, for example. The selection unit 13 compares each of these predictions with the actual values ​​for FY20 and selects the model with the smallest prediction error rate.

[0026] In the production planning process S16, the production planning unit 14 uses the model selected by the selection unit 13 to make a prediction for fiscal year 21 based on past performance data from the 10 years from fiscal year 2011 to fiscal year 2020, and prepares a production plan for fiscal year 21.

[0027] During production, manufacturers carry out daily production based on a production plan, but because order volumes change daily, there are times when the production plan is not enough to meet demand. In such cases, manufacturers respond to the fluctuating order volumes by adjusting the number of units produced each day as needed. Since production adjustments have a negative impact on material procurement and inventory, it is important to create a production plan that minimizes production adjustments. In other words, it is important to select an appropriate forecasting model.

[0028] In forecasts for each model, the number of units produced per model is small, so there is a large variation. In order to reduce the variation, it is better to broaden the classification from small classification prediction to large classification prediction and make predictions in units of increased number of units, which makes it possible to make stable predictions and improves forecast accuracy.

[0029] When only a specific model group shows characteristic secular changes, forecasting using only that model group will improve forecast accuracy. For example, when economic activity declines across society as a whole, such as during the Lehman Shock, order volumes will show similar changes across many models, so a broad category forecast or overall forecast will yield more accurate forecasts than a narrow category forecast.

[0030] In the current COVID-19 pandemic, economic activity declines across society as a whole, but in particular, orders for ventilation fans are increasing for only certain models in order to improve ventilation, and different trends may be observed depending on the model group.

[0031] In other words, whether the accuracy is higher for major classifications or for minor classifications varies depending on the social situation. The production planning system 1 according to the first embodiment selects a forecasting method with high forecast accuracy from among multiple size classification forecasts at the time of forecasting demand and determines the number of units to be produced, so that predictions can be made that correspond to fluctuations in economic activity, thereby improving forecast accuracy. In other words, the production planning system 1 can select an appropriate demand forecasting model.

[0032] Embodiment 2 In embodiment 2, a production planning system will be described in which the machine learning prediction unit 25 included in the production planning system 1 according to embodiment 1 exerts its functions. As described in embodiment 1, the machine learning prediction unit 25 predicts demand by machine learning.

[0033] The target product A is a production planning system. The input 2 (state) B2 is past performance data. The input 1 (action) B1 is a method for predicting demand from past performance data. Data related to demand prediction includes the period of past data to be referenced, model classification, aging change estimation method, monthly fluctuation estimation method, and model-specific allocation estimation method. The output C is a production plan. Of the remuneration criteria D, the remuneration increase criterion D1 corresponds to a decrease in the prediction error rate. The remuneration decrease criterion D2 corresponds to an increase in the prediction error rate.

[0034] In the second embodiment, similar to the first embodiment, a case will be considered in which evaluation and prediction for fiscal year 20 is performed using data from fiscal year 2019 and before. In this case, B2 input 2 (state) is actual data from fiscal year 2018 and before.

[0035] B1 shows details of input 1 (behavior).

[0036] First, the machine learning prediction unit 25 selects, as a variable, the number of years of the period to be used as a reference period from the past data. For example, if the reference period is 10 years, 10 is selected as the variable.

[0037] Next, the machine learning prediction unit 25 selects a model classification method. Model classification methods include model-specific classification, minor classification, major classification, and overall classification, as described in the first embodiment, but other methods are also possible. The machine learning prediction unit 25 can select any combination.

[0038] Next, the machine learning prediction unit 25 selects a secular change estimation method. The approximation method includes linear approximation, exponential approximation, logarithmic approximation, and polynomial approximation using the time axis as a variable. The degree of polynomial approximation can be selected up to (reference years - 1).

[0039] Next, the machine learning prediction unit 25 selects a monthly fluctuation estimation method. As the approximation method for monthly fluctuation estimation, constant approximation, linear approximation, exponential approximation, logarithmic approximation, or polynomial approximation can be selected.

[0040] Next, the machine learning prediction unit 25 selects a method for estimating the allocation for each model. As the approximation method for the allocation for each model, constant approximation, linear approximation, exponential approximation, logarithmic approximation, or polynomial approximation can be selected.

[0041] B1 input 1 (action) is determined by selecting the reference period, model classification method, secular change estimation method, monthly fluctuation estimation method, and model-specific allocation estimation method, and the machine learning prediction unit 25 predicts demand for FY2019, the year subject to machine learning evaluation.

[0042] In this way, the machine learning prediction unit 25 compares the selected conditions with the performance data for fiscal year 2019 predicted based on performance data for fiscal year 2018 and earlier, and determines the reward standard based on the difference from the performance data. In other words, the machine learning prediction unit 25 increases the reward when the conditions are such that the error is small.

[0043] The machine learning prediction unit 25 changes the condition selection and repeats the reward judgment, thereby learning the optimal behavior, that is, the optimal condition selection.

[0044] Details of the machine learning are given below.

[0045] Learning Phase FIG. 8 is a diagram showing the configuration of the learning device 2 according to the second embodiment. FIG. 8 shows the configuration of the learning device 2 for target product A. The learning device 2 has a data acquisition unit 3 and a model generation unit 4. The data acquisition unit 3 acquires learning data including past performance data of the production planning system and a demand forecast of the production planning system based on the past performance data. The model generation unit 4 uses the learning data acquired by the data acquisition unit 3 to generate a trained model for inferring a production plan from the past performance data of the production planning system. The learning device 2 will be further described below.

[0046] The data acquisition unit 3 acquires B1 input 1 (action) and B2 input 2 (state) as learning data.

[0047] The model generation unit 4 learns the C output based on the learning data including the B1 input 1 (action) and the B2 input 2 (state). That is, the model generation unit 4 generates a learned model that infers the C output from the B2 input 2 (state) of the A target product.

[0048] The learning algorithm used by the model generation unit 4 can be a known algorithm such as supervised learning, unsupervised learning, or reinforcement learning. As an example, a case where reinforcement learning is applied to the learning algorithm used by the model generation unit 4 will be described. In reinforcement learning, an agent (subject of action) in a certain environment observes the current state (environmental parameters) and decides on an action to be taken. The environment changes dynamically depending on the agent's actions, and the agent is given a reward according to the change in the environment. The agent repeats this process and learns the course of action that will obtain the most reward through a series of actions. Q-learning and TD (Temporal Difference) learning are known as representative methods of reinforcement learning. For example, in the case of Q-learning, an action value function Q(s t , a t A general update formula for the parameter θ is expressed by the following formula (1):

[0049]

[0050] In formula (1), s t represents the state of the environment at time t, and a t represents the action at time t. t Therefore, the state is s t+1 It changes to r t+1 represents the reward that can be obtained depending on the change in state, γ represents the discount rate, and α represents the learning coefficient. Note that γ is in the range of 0<γ≦1, and α is in the range of 0<α≦1. B1 Input 1 (action) is action a t and B2 input 2 (state) is state s t The learning device 2 is in the state s at time t. t Best action in a t Learn.

[0051] The update formula expressed by formula (1) increases the action value Q if the action value Q of the action a with the highest Q value at time t+1 is greater than the action value Q of the action a executed at time t, and decreases the action value Q in the opposite case. In other words, the learning device 2 updates the action value function Q(s) so that the action value Q of the action a at time t approaches the best action value at time t+1. t , a t) is updated. This allows the best action value in a certain environment to be propagated to the action value in the previous environment.

[0052] As described above, when generating a trained model by reinforcement learning, the model generation unit 4 has a reward calculation unit 41 and a function update unit 42.

[0053] The reward calculation unit 41 calculates a reward based on the B1 input 1 (action) and the B2 input 2 (state). The reward calculation unit 41 calculates a reward r based on the D reward criterion (a collective term for D1 and D2). For example, the reward calculation unit 41 increases the reward r (for example, gives a reward of "1") in the case of the D1 reward increase criterion, and decreases the reward r (for example, gives a reward of "-1") in the case of the D2 reward decrease criterion.

[0054] The function update unit 42 updates the function for determining the C output in accordance with the reward calculated by the reward calculation unit 41, and outputs the C output to the trained model storage unit 5. For example, in the case of Q-learning, the function update unit 42 updates the action value function Q(s t , a t ) is used as a function to calculate the C output.

[0055] The learning device 2 repeatedly executes the above-described learning. The learned model storage unit 5 stores the action value function Q(s t , a t ), that is, stores the trained model. The trained model storage unit 5 is realized by, for example, a semiconductor memory.

[0056] Next, the learning process performed by the learning device 2 will be described with reference to Fig. 9. Fig. 9 is a flowchart showing the procedure of the operation related to the learning process performed by the learning device 2 according to the second embodiment.

[0057] In step S21, the data acquisition unit 3 acquires B1 input 1 (action) and B2 input 2 (state) as learning data.

[0058] In step S22, the model generation unit 4 calculates a reward based on the B1 input 1 (action) and the B2 input 2 (state). Specifically, the reward calculation unit 41 acquires the B1 input 1 (action) and the B2 input 2 (state) and determines whether to increase or decrease the reward based on a predetermined D reward standard.

[0059] If the remuneration calculation unit 41 determines that the remuneration should be increased (determining whether to increase the remuneration in step S22), it increases the remuneration in step S23. If the remuneration calculation unit 41 determines that the remuneration should be decreased (determining whether to decrease the remuneration in step S22), it decreases the remuneration in step S24.

[0060] In step S25, the function update unit 42 updates the action value function Q(s) represented by equation (1) stored in the trained model storage unit 5 based on the reward calculated by the reward calculation unit 41. t , a t ) to update.

[0061] The learning device 2 repeatedly executes the operations from step S21 to step S25 and generates the action value function Q(s t , a t ) is stored as a learned model in the learned model storage unit 5.

[0062] The learning device 2 of embodiment 2 stores the learned model in a learned model memory unit 5 provided outside the learning device 2, but the learned model memory unit 5 may also be provided inside the learning device 2.

[0063] Utilization Phase FIG. 10 is a diagram showing the configuration of the inference device 6 according to the second embodiment. FIG. 10 shows the configuration of the inference device 6 for target product A. The inference device 6 has a data acquisition unit 7 and an inference unit 8. The data acquisition unit 7 acquires past performance data of the production planning system. The inference unit 8 outputs a production plan from the past performance data acquired by the data acquisition unit 7, using a trained model for inferring a demand forecast for the production planning system from the past performance data acquired by the data acquisition unit 7. The inference device 6 is further described below.

[0064] The data acquisition unit 7 acquires the B2 input 2 (state).

[0065] The inference unit 8 infers the C output using the trained model stored in the trained model storage unit 5. That is, the inference unit 8 can infer the C output suitable for the B2 input 2 (state) by inputting the B2 input 2 (state) acquired by the data acquisition unit 7 into the trained model.

[0066] In the second embodiment, the inference unit 8 outputs the C output using the trained model trained by the model generation unit 4 of the A target product, but it may also acquire a trained model from another A target product and output the C output based on this trained model.

[0067] Next, the process by which the inference device 6 obtains the C output will be described with reference to Fig. 11. Fig. 11 is a flowchart showing the procedure of the operation performed by the inference device 6 and the production planning system according to the second embodiment.

[0068] In step S31, the data acquisition unit 7 acquires the B2 input 2 (state).

[0069] In step S32, the inference unit 8 inputs B2 input 2 (state) to the learned model stored in the learned model storage unit 5 and obtains C output.

[0070] In step S33, the inference unit 8 outputs the C output obtained by the trained model to the A target product.

[0071] In step S34, the production planning system for target product A predicts product demand and prepares a production plan using the production plan that is output as output C. This enables the production planning system to prepare a production plan with a small error from the actual production and with little production adjustment.

[0072] In the second embodiment, a case has been described in which reinforcement learning is applied to the learning algorithm used by the inference unit 8, but the present invention is not limited to this. As for the learning algorithm, it is also possible to apply supervised learning, unsupervised learning, semi-supervised learning, or the like, in addition to reinforcement learning.

[0073] Furthermore, the learning algorithm used by the model generation unit 4 can be deep learning, which learns to extract the features themselves, or the learning device 2 can perform machine learning according to other known methods, such as neural networks, genetic programming, functional logic programming, or support vector machines.

[0074] The learning device 2 and the inference device 6 may be connected to target product A via a network and may be separate devices from target product A. Alternatively, the learning device 2 and the inference device 6 may be built into target product A. Furthermore, the learning device 2 and the inference device 6 may exist on a cloud server.

[0075] Furthermore, the model generation unit 4 may learn the C output using learning data acquired from multiple target products A. The model generation unit 4 may acquire learning data from multiple target products A used in the same area, or may learn the C output using learning data collected from multiple target products A operating independently in different areas. It is also possible to add or remove target products A from which learning data is collected as targets during the process. Furthermore, a learning device that has learned the C output for a certain target product A may be applied to another target product A, and the learning device may re-learn the C output for the other target product A to generate the action value function Q(s t , a t ) may be updated.

[0076] Fig. 12 is a flowchart showing the procedure of the overall operation of the production plan according to embodiment 2. Fig. 12 is a flowchart in which a machine learning prediction step S50 is added to the overall operation of the production plan according to embodiment 1. Fig. 13 is a schematic diagram of evaluation prediction according to embodiment 2. Fig. 14 is a schematic diagram of planning according to embodiment 2.

[0077] In the model-specific prediction step S11, the small category prediction step S12, the large category prediction step S13, and the overall prediction step S14, the production planning system uses data from the year prior to the evaluation prediction year and makes an evaluation prediction for the evaluation prediction year in the same manner as in embodiment 1.

[0078] For example, if the production planning system performs evaluation and prediction in fiscal year 20 to create a production plan for fiscal year 21, the production planning system performs evaluation and prediction for fiscal year 20 using actual performance data up to fiscal year 19.

[0079] In the machine learning prediction step S50, the machine learning prediction unit 25 uses a trained model that has been trained by comparing predicted values ​​learned using actual data from the year before the year before the evaluation prediction year with actual data from the previous year, to make an evaluation prediction for the evaluation prediction year using data from the year before the evaluation prediction year and earlier.

[0080] If the machine learning prediction unit 25 makes an evaluation prediction in fiscal year 2020 in order to make a production plan for fiscal year 21, it creates a trained model by comparing the learned prediction based on actual data up to fiscal year 2018 with the actual results for fiscal year 2019, and makes an evaluation prediction for fiscal year 2020 using the created trained model and the actual data up to fiscal year 2019.

[0081] In the selection step S15, the production planning system selects the model with the smallest prediction error rate over the past year from among the model-specific prediction, small category prediction, large category prediction, overall prediction, and machine learning prediction.

[0082] In the production planning step S16, the production planning system uses the selected model and actual performance data up to the evaluation forecast year, i.e., up to the year before the planned year, to carry out a planning forecast for the planning year, which is the year following the evaluation forecast year, and creates a production plan for the planning year.

[0083] In this way, the production planning system according to the second embodiment can compare the validity of the machine learning prediction with other predictions and select the prediction model appropriately. Furthermore, the production planning system according to the second embodiment determines the remuneration standard based on the difference between the demand forecast for the previous year, which is predicted based on the actual data up to the year before the evaluation forecast year, and the actual data for the previous year.

[0084] 15 is a diagram showing a processor 97 in a case where some or all of the functions of the performance data acquisition unit 11, the prediction unit 12, the selection unit 13, and the production planning unit 14 of the production planning system 1 according to the first embodiment are realized by the processor 97. In other words, some or all of the functions of the performance data acquisition unit 11, the prediction unit 12, the selection unit 13, and the production planning unit 14 may be realized by the processor 97 that executes a program stored in a memory 98. The processor 97 is a CPU (Central Processing Unit), a processing system, an arithmetic system, a microprocessor, or a DSP (Digital Signal Processor). The memory 98 is also shown in FIG.

[0085] When some or all of the functions of the performance data acquisition unit 11, the prediction unit 12, the selection unit 13, and the production planning unit 14 are realized by the processor 97, the functions are realized by the processor 97 and software, firmware, or a combination of software and firmware. The software or firmware is written as a program and stored in the memory 98. The processor 97 reads and executes the program stored in the memory 98, thereby realizing some or all of the functions of the performance data acquisition unit 11, the prediction unit 12, the selection unit 13, and the production planning unit 14.

[0086] When some or all of the functions of the performance data acquisition unit 11, the prediction unit 12, the selection unit 13, and the production planning unit 14 are realized by the processor 97, the production planning system 1 has a memory 98 for storing a program that results in the execution of some or all of the steps executed by the performance data acquisition unit 11, the prediction unit 12, the selection unit 13, and the production planning unit 14. It can also be said that the program stored in the memory 98 causes a computer to execute some or all of the procedures or methods executed by the performance data acquisition unit 11, the prediction unit 12, the selection unit 13, and the production planning unit 14.

[0087] The memory 98 may be, for example, a non-volatile or volatile semiconductor memory such as a RAM (Random Access Memory), a ROM (Read Only Memory), a flash memory, an EPROM (Erasable Programmable Read Only Memory), an EEPROM (Electrically Erasable Programmable Read-Only Memory), a magnetic disk, a flexible disk, an optical disk, a compact disk, a mini disk, or a DVD (Digital Versatile Disk).

[0088] 16 is a diagram showing a processing circuit 99 in a case where some or all of the functions of the performance data acquisition unit 11, the prediction unit 12, the selection unit 13, and the production planning unit 14 of the production planning system 1 according to the first embodiment are realized by the processing circuit 99. In other words, some or all of the performance data acquisition unit 11, the prediction unit 12, the selection unit 13, and the production planning unit 14 may be realized by the processing circuit 99.

[0089] The processing circuitry 99 is dedicated hardware, such as a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination thereof.

[0090] Some of the functions of the performance data acquisition unit 11, the prediction unit 12, the selection unit 13, and the production planning unit 14 may be realized by dedicated hardware separate from the hardware that realizes the remaining functions of the performance data acquisition unit 11, the prediction unit 12, the selection unit 13, and the production planning unit 14.

[0091] With regard to the multiple functions of the performance data acquisition unit 11, the prediction unit 12, the selection unit 13, and the production planning unit 14, some of the multiple functions may be realized by software or firmware, and the remaining multiple functions may be realized by dedicated hardware. In this way, the multiple functions of the performance data acquisition unit 11, the prediction unit 12, the selection unit 13, and the production planning unit 14 can be realized by hardware, software, firmware, or a combination of these.

[0092] As described above, the prediction unit 12 includes the overall prediction unit 21, the major classification prediction unit 22, the minor classification prediction unit 23, the model-specific prediction unit 24, and the machine learning prediction unit 25. Some or all of the functions of the overall prediction unit 21, the major classification prediction unit 22, the minor classification prediction unit 23, the model-specific prediction unit 24, and the machine learning prediction unit 25 may be implemented by a processor or a processing circuit. The processor is a processor equivalent to the processor 97. The processing circuit is a processing circuit equivalent to the processing circuit 99.

[0093] Some or all of the functions of the data acquisition unit 3 and the model generation unit 4 of the learning device 2 according to the second embodiment may be realized by a processor or a processing circuit. The processor is a processor equivalent to the processor 97. The processing circuit is a processing circuit equivalent to the processing circuit 99. Some or all of the functions of the reward calculation unit 41 and the function update unit 42 of the model generation unit 4 may be realized by a processor or a processing circuit. Furthermore, each of the reward calculation unit 41 and the function update unit 42 may be realized by a processing circuit.

[0094] Some or all of the functions of the data acquisition unit 7 and the inference unit 8 of the inference device 6 according to the second embodiment may be realized by a processor or a processing circuit. The processor is a processor equivalent to the processor 97. The processing circuit is a processing circuit equivalent to the processing circuit 99.

[0095] The configurations shown in the above embodiments are merely examples, and may be combined with other known technologies, or different embodiments may be combined with each other, or part of the configuration may be omitted or modified within the scope of the gist of the invention.

[0096] 1 Production planning system, 2 Learning device, 3 Data acquisition unit, 4 Model generation unit, 5 Learned model storage unit, 6 Inference device, 7 Data acquisition unit, 8 Inference unit, 11 Performance data acquisition unit, 12 Prediction unit, 13 Selection unit, 14 Production planning unit, 21 Overall prediction unit, 22 Major classification prediction unit, 23 Minor classification prediction unit, 24 Model-specific prediction unit, 25 Machine learning prediction unit, 41 Reward calculation unit, 42 Function update unit, 97 Processor, 98 Memory, 99 Processing circuit.

Claims

1. A production planning system which plans the production volume of a plurality of products in a predetermined unit period, comprising: a performance data acquisition unit which acquires past sales performance data for each of the plurality of products; a model-by-model prediction unit which predicts demand for each product model based on the sales performance data acquired by the performance data acquisition unit; a subcategory prediction unit which classifies products into subcategories including a plurality of models and predicts demand for each subcategory based on the sales performance data acquired by the performance data acquisition unit; an overall prediction unit which predicts demand for all products based on the sales performance data acquired by the performance data acquisition unit; a selection unit which selects an optimal prediction model which has the minimum prediction error rate for the past year from among the demand forecasts predicted by a prediction unit having the model-by-model prediction unit, the subcategory prediction unit, and the overall prediction unit; and a production planning unit which formulates a production plan based on the optimal prediction model selected by the selection unit.

2. The production planning system according to claim 1, further comprising a major classification forecasting unit which classifies products into major classifications including a plurality of minor classifications based on the sales performance data acquired by the performance data acquisition unit, and forecasts demand for each major classification.

3. The production planning system according to claim 1, characterized in that the forecasting unit forecasts demand based on a forecast value estimated based on the change in annual sales performance values ​​over time and a seasonal fluctuation value which is the difference between the average sales performance value and the average performance value for a unit period.

4. The production planning system according to claim 1, further comprising a machine learning prediction unit that predicts demand through machine learning.

5. The production planning system according to claim 4, characterized in that the remuneration standard is determined based on the difference between the demand forecast for the previous year, which is predicted based on actual data from the year before the evaluation forecast year, and the actual data for the previous year.

6. A learning device comprising: a data acquisition unit that acquires learning data including past performance data of a production planning system and a demand forecast of the production planning system based on the past performance data; and a model generation unit that uses the learning data acquired by the data acquisition unit to generate a trained model for inferring a production plan from the past performance data of the production planning system.

7. An inference device comprising: a data acquisition unit that acquires past performance data of a production planning system; and an inference unit that outputs a production plan from the past performance data acquired by the data acquisition unit using a trained model for inferring a demand forecast of the production planning system from the past performance data acquired by the data acquisition unit.

Citation Information

Patent Citations

  • Machine learning device, production plan formulation device and inference device

    JP2023013058A

  • Learning data generation device, program and learning data generation method

    JP2023031427A

  • Economic indicator prediction device, production planning device, prediction model generation device, and prediction model generation method

    WO2021059480A1