KR102574244B1 - A method and apparatus for generating future demand forecast data based on attention mechanism - Google Patents

Abstract

The present invention proposes a method and apparatus for generating future demand forecast data based on an attention mechanism capable of improving the accuracy of future demand forecast for each product that may occur in industrial settings. The present invention is a configuration including a data preparation engine unit that prepares data for forecasting demand at a future time point using past demand data, and a prediction/analysis engine unit that predicts and analyzes demand at a future time point based on the prepared data, wherein The data preparation engine unit converts the past demand data into time-series data so that there are no missing values when there is an intermittent pattern, and the prediction/analysis engine unit decomposes the future forecast data for intermittent demand into individual future forecast data using a neural network model. and provide

Description

A method and apparatus for generating future demand forecast data based on attention mechanism {A method and apparatus for generating future demand forecast data based on attention mechanism}

The present invention accurately predicts demand at a future point in time by using a Recurrent Neural Network (RNN) neural network algorithm to which an Attention Mechanism is applied regardless of the data type of time series data or intermittent data generated at industrial sites. It relates to a method and apparatus capable of generating data.

Demand forecasting for the future in all fields is directly related to corporate profits, and it is very important for decision-making in sectors including product production and marketing as well as business strategies for efficient supply chain management. Accurate demand forecasting enables efficient operation of companies, and through this, companies can provide satisfactory services to customers.

However, inaccurate forecasting can lead to overstocking, low service levels, and inefficient use of resources, affecting overall supply chain operations. So, forecasting the demand for a product is very important. In particular, it can be an important issue for small and medium-sized manufacturers, as they have difficulties in market research or demand forecasting, resulting in large operating losses due to product inventory.

It is difficult for small and medium-sized manufacturers to establish a mass production system for products, so MTO (Make To Order) provides products in case of quotation and actual demand for a specific product by multi-product production for small-quantity products, that is, multi-product small-lot production method. There are cases in which a production system is operated. This multi-variety, small-volume production system is characterized by relatively high dependency on orders from customers, that is, a relatively high impact on the production system due to changes in demand received. However, these characteristics cause problems such as instability in production speed, frequent scheduling disruptions, and difficulties in procuring raw materials on time and in the right quantity. In other words, it is not possible to prepare stocks of materials such as raw materials and semi-finished products in an appropriate quantity in advance, so a fatal problem of failing to place an order within the deadline can occur and can be expected. That is why forecasting demand is so important. Therefore, it is important to accurately predict the intermittent demand that appears in the multi-product, small-volume production system.

Conventional demand forecasting techniques include ARIMA and exponential smoothing. However, these techniques are time-series data prediction techniques, which provide sufficient performance for continuous data where there is a continuous demand, that is, there are not many missing values. It falls. Intermittent demand data means a case in which the demand for a specific product that occurred in the past does not occur regularly in a time unit defined by the system (eg, perimeter, month, quarter, etc.).

Korea Patent Publication No. 10-2005-0104482 (2005. 11. 03. Production planning system and method through automatic demand forecasting)

Accordingly, an object of the present invention is to provide a method and apparatus for generating future demand forecast data based on an attention mechanism capable of improving the accuracy of future demand forecast for each product that may occur in a series of industrial settings.

Another object of the present invention is to enable accurate demand forecasting even when product demand occurs intermittently.

That is, the present invention will generate, analyze, and provide accurate demand forecast data for the future using an attention mechanism-based RNN neural network (ie, RNN Attention) regardless of time series data or intermittent data.

The present invention for achieving the above object, the data preparation engine unit for preparing the data for demand forecast of the future point using the past demand data; A prediction/analysis engine unit for predicting and analyzing demand at a future point in time based on prepared data, wherein the data preparation engine unit converts past demand data into time-series data so that there is no missing information when the demand data shows an intermittent pattern, and the prediction/analysis engine Part 2 provides a future demand forecasting data generation device characterized in that future forecasting data for intermittent demand is decomposed into individual future forecasting data using a neural network model and provided.

The data preparation engine unit may include an account data generation unit that sets a database range to be used for demand forecasting and creates an account according to an aggregation standard; A pre-processing unit that pre-processes accounts at all times; and an input information providing unit that collects past demand data according to the preprocessed value, generates a smart set converted into time series data, and provides the generated smart set to the prediction/analysis engine unit.

The neural network model is based on a Recurrent Neural Network (RNN) using an Attention Mechanism.

The prediction/analysis engine unit includes a neural network model providing unit, and the neural network model providing unit includes: a neural network model execution unit receiving the smart set as an input and learning; a first calculation unit to calculate weights for the smart set; a function unit that equalizes the weights; a second calculation unit calculating a homogenized smart set using the smart set and a homogenized weight; and a third calculation unit that calculates individual weights for each product using the homogenized smart set, and provides a future demand forecast value for each product through the individual weights for each product.

According to another feature of the present invention, an apparatus for generating demand forecast data at a future point in time using past demand data, comprising: converting intermittent demand pattern data into time-series data; learning the time series data with a neural network; and providing individual future demand forecast data for products according to learning results.

The method further includes providing an analysis report on which past pattern has been derived using the individual future demand forecast data.

The learning of the neural network may include calculating a smart set by adding past demand data generated for each product at each point in time; inputting the calculated smart set information to the neural network; Calculating a weight for each product according to learning of the neural network; equalizing time-specific weights using a predetermined function; calculating a homogenized smart set by calculating a smart set and a homogenization weight for each product; and calculating individual weights for each product at each point in time using the homogenized smart set, and generating a predicted future value for each product.

In the step of learning the neural network, an Attention Mechanism-based Recurrent Neural Network (RNN) neural network is used.

The function applies the softmax function.

According to the attention mechanism-based future demand forecast data generation method and apparatus of the present invention as described above, while applying the attention mechanism-based RNN neural network model, in the case of intermittent pattern demand data, after converting it into a predictable time series pattern, individual products Since it is possible to provide a future forecast value for a specific product, even when demand data occurs in a time-series pattern or an intermittent pattern for a given product, there is an effect of accurately predicting demand at a specific future point in time. Therefore, in the case of products sold in most medium/small businesses, it is frequently difficult to apply the conventional demand forecasting technique, so more improved forecasting results can be expected.

According to the present invention, the future demand forecast value provides analysis data on how much a factor or time period acted in the past demand pattern through numerical values and visualization, so that the forecast result value can be approximated to a specific target value. You can more effectively establish management strategies, such as analyzing time zones.

1 is an overall configuration diagram of an apparatus for generating future demand forecast data based on an attention mechanism according to an embodiment of the present invention.
Figure 2 is a block diagram of the neural network model provider shown in Figure 1
3 is a flowchart explaining a method for generating future demand forecast data according to an embodiment of the present invention.
Figure 4 is a flow chart explaining the execution process of the neural network model according to the present invention
5 is a table for numerically explaining the process of FIG. 4

Since the present invention can apply various transformations and have various embodiments, specific embodiments will be illustrated in the drawings and described in detail. However, it should be understood that this is not intended to limit the specific embodiments of the present invention, and includes all conversions, equivalents, and substitutes included in the spirit and scope of the present invention. In describing the present invention, if it is determined that a detailed description of related known technologies may obscure the gist of the present invention, the detailed description will be omitted.

Terms such as first and second may be used to describe various components, but the components should not be limited by the terms. These terms are only used for the purpose of distinguishing one component from another.

Terms used in the present invention are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly dictates otherwise. In this application, the terms "include" or "have" are intended to designate that there is a feature, number, step, operation, component, part, or combination thereof described in the specification, but one or more other features It should be understood that the presence or addition of numbers, steps, operations, components, parts, or combinations thereof is not precluded.

Spatially relative terms, such as below, beneath, lower, above, upper, etc., facilitate the correlation between one element or component and another element or component, as shown in the drawing. can be used to describe Spatially relative terms should be understood as encompassing different orientations of elements in use or operation in addition to the orientations shown in the figures. For example, when an element shown in the drawing is turned over, an element described as below or beneath another element may be placed above or above the other element. Accordingly, the exemplary term below may include both directions of down and above. Elements may also be oriented in other orientations, and thus spatially relative terms may be interpreted according to orientation.

As used in the present invention, an expression indicating a part such as “part” or “part” refers to a device in which a corresponding component may include a specific function, software which may include a specific function, or a device which may include a specific function. and software, but cannot necessarily be limited to the expressed functions, which are provided only to help a more general understanding of the present invention, and those with ordinary knowledge in the field to which the present invention belongs If so, various modifications and variations are possible from these descriptions.

In addition, it should be noted that all electrical signals used in the present invention, as an example, can be reversed in signs of all electrical signals to be described below when an inverter or the like is additionally provided in the circuit of the present invention. Therefore, the scope of the present invention is not limited to the direction of the signal.

Therefore, the spirit of the present invention should not be limited to the described embodiments, and it will be said that not only the claims to be described later, but also all modifications equivalent or equivalent to these claims fall within the scope of the spirit of the present invention. .

Hereinafter, the present invention will be described in more detail based on the embodiments shown in the drawings.

1 is an overall configuration diagram of an apparatus for generating future demand prediction data based on an attention mechanism according to an embodiment of the present invention. As shown, the apparatus 10 of the present invention largely includes a data preparation engine unit 100 that prepares data and a prediction/analysis engine unit 200 that performs a prediction and analysis process based on the prepared data.

First, the data preparation engine unit 100 includes an account data generation unit 110, a pre-processing unit 120, and an input information calculation unit 130. The account data generation unit 110 is a component that sets the range of the database to be used for demand forecasting and performs a processor that generates an account, which is individual time series data, according to aggregation criteria such as product name/ordering place. A data range may mean selecting a criterion, for example in units of time. A total of 12 time series data are created when the data range is selected by month based on a year, and a total of 52 time series data are created when selected by week.

The pre-processing unit 120 provides a process of pre-processing the created account. The pre-processing targets include basic value processing for setting basic information (e.g., currency setting such as won and dollar, prediction algorithm allocation, etc.), integrated account (smart set) processing, outlier processing that selects whether or not to remove data showing outliers using a smoothing method, and target value processing, which is the standard for predicted target values such as actual demand and actual sales.

The input information providing unit 130 collects and refines data according to the preprocessed values, and then calculates and provides input information that can be used by the prediction/analysis engine unit 200 at a later stage.

Second, the prediction/analysis engine unit 200 is configured to predict and analyze future demand values by learning the input information provided by the data preparation engine unit 100 in the neural network model, and includes the model fitting unit 210 and the neural network model. It includes a providing unit 220, a result providing unit 230, and a correcting unit 240.

The model fitting unit 210 provides a process for selecting parameters for the prediction algorithm of the neural network. For example, a parameter with the smallest error value compared to the amount of learning by the prediction algorithm can be selected. The neural network model providing unit 220 provides a process of generating future demand prediction data based on past demand data. The result providing unit 230 provides a process for providing future prediction data so that the user can check it, and may provide visualized data to easily check the future prediction data if necessary. The correction unit 240 provides a process of correcting future demand forecast data. If future demand forecast data needs to be additionally provided, or if it is generated incorrectly or if unsatisfactory result values are provided, it is re-learned and reproduced.

The neural network model of the present invention is based on a Recurrent Neural Network (RNN). The RNN-based neural network produces and provides future demand forecast data through the Attention Mechanism. The attention mechanism operates in the order of attention score, attention output, attention distribution, and decoder hidden state, and as a known configuration, a detailed description thereof will be omitted in this specification.

FIG. 2 is a block diagram of the neural network model providing unit shown in FIG. 1 . As shown, the neural network model provider 220 includes a smart set calculator 221 that calculates a smart set by adding past demand data generated for each product at each point in time, and a first calculator that calculates a weight for the smart set ( 222), a function unit 223 that equalizes the weights, a second calculator 224 that calculates a uniformed smart set using the smart set and the uniformed weights, and individual weights for each product using the uniformed smart set It is configured to include a third calculator 225 that calculates. According to this configuration, a future demand forecast value for each product is provided through the individual weights for each product.

Next, a process of generating and providing future demand forecasting data using the future demand forecasting data generator constructed as described above will be described.

3 is a flowchart illustrating a method for generating future demand forecast data according to an embodiment of the present invention. In the present invention, the step of preparing data for forecasting future demand (data preparation step, s100), and the step of predicting and analyzing demand at a future point based on the prepared data (prediction/analysis step, s200)) proceed sequentially do.

In the data preparation step (s100), first, the range of the entire database to be used for demand forecasting is set (s110). The criterion for clustering the past demand data is selected (s120), and an account is created according to each criterion (s130). For example, if the past demand data is based on the product group, it can be classified as product 1 to product n, and an account is created for each product group. At this time, the account may be a normal demand pattern in which demand continuously occurs or an intermittent demand pattern. This may be different for each product group.

When accounts are created according to the cluster criteria, policy management information for the created accounts is set (s140). Policy management refers to setting the pre-processing unit 120 to predict future demand according to which criterion based on past demand data. Examples of policy management include basic information setting, smart set processing, outlier management, target value setting, and the like. Setting basic information is, for example, setting a currency value such as won (\) or dollar ($), and setting a prediction algorithm. Smart set processing is an operation of integrating into one account, a smart set, when the created accounts show intermittent demand. If there is intermittent demand, it can be regarded as setting to do a smart set. Outlier management is to select whether or not to remove data using a smoothing method when there is data that appears to be an outlier. Target value setting is to set standards for forecast target values, such as actual demand or actual sales.

Thereafter, the input information providing unit 130 collects and refines data for forecasting future demand, and then processes the data into a form that can be used for prediction and analysis (s150). The input information providing unit 130 integrates accounts with missing values and converts them into one time series data (ie, smart set) to create a new data set. The generated data set is actually used in the future prediction step. The reason for converting to time series data here is that it is difficult to accurately predict future demand if the past demand data for forecasting future demand is an intermittent demand pattern. An intermittent demand pattern refers to data in which a number of missing values occur in the time unit defined by the system for the demand for a specific product that occurred in the past. Therefore, it can be considered that the input information providing unit 130 performs an operation of converting the demand data generated in the past into a time-series data form in which missing values are not visible. Of course, if the past demand data is a normal pattern rather than an intermittent demand pattern, the above conversion operation may be omitted.

Through this process, past demand data is processed into a data set having a time-series data pattern capable of predicting actual future demand, and future demand forecasting will be performed using the processed data set. We continue our look at demand forecasting and analysis.

When a data set is received, the prediction/analysis engine unit 200 performs a model fitting process of learning time-series data from a predetermined neural network model (S210). This process may be the task of determining the parameters by which the neural network model can optimally perform future demand forecasting according to learning.

Thereafter, the neural network model provided by the neural network model providing unit 220 is executed (S220) to predict future demand for the data set, and a future demand forecast value is calculated through the RNN-based neural network model (S230). A detailed description of step 220 will be described below with reference to FIG. 4 . And the calculated future demand forecast value is provided so that the user can check it by the result providing unit (s240). The calculated predicted result value may be provided through visualization as well as numerical values. Here, since the present invention provides weights through the attention mechanism, it is possible to provide an explanation through figures and visualizations for future demand prediction results, and the factors of past demand data acting on the calculated future demand prediction results or time period, and contribution can be known. In other words, it is possible to provide interpretive power that can explain the basis on which demand forecast result value was predicted (s250).

Therefore, it is possible to analyze related factors and time periods so that the future demand forecast value can approach the specific target value described above, so that the company's management strategy can be established more effectively.

In the present invention, a process of correcting the predicted future demand may be selectively performed. A correction process may be performed when a future demand forecast is to be practically re-examined, such as a future demand forecast based on a change in reference value or an error in past demand data or data set.

4 is a flowchart illustrating an execution process of a neural network model for predicting future demand according to the present invention. This process is performed by the neural network model providing unit 220 .

As shown, a product whose demand is to be predicted is set (s300). The data for the demand forecasting product has been completed in advance in the data preparation step (s100) described above, and here, the data set corresponding to the demand forecasting target product can be called. The data set may be normal pattern data or intermittent demand pattern data. Normal pattern data is data without a large number of missing values, and may be data that does not require an operation of generating a new data set by converting it into the above-mentioned time series data.

When the data set is input, the neural network model starts learning to calculate a predicted value (S320). According to the learning process, a process of determining a past demand pattern is performed (S330). In the above step, in the case of a data set of normal pattern data, the neural network model calculates and provides time weights and future predicted values for each product (S340). Through this, it is possible to know from which time series pattern in the past the future demand forecast has been made.

On the other hand, even in the case of a data set having intermittent pattern data (s350), time weights and future predicted values may be calculated and provided for each product. However, in the case of an intermittent demand pattern, since it is not known which data pattern was influenced by the past, it is necessary to allocate future demand forecasts by product again.

In order to calculate an accurate predicted value, the smart set calculation unit 221 calculates a smart set by adding past demand data generated for each product at each point in time (S352). Then, the calculation unit 222 calculates the time weight for each product (S354), and the function unit 223 to which softmax is applied equalizes the calculated time weight for each product and time point (S354). The uniformized time weight for each product is referred to as 'uniform weight'. When all of the equalization weights are added, the sum becomes '1'.

The second calculation unit 224 recalculates the smart set s' using the data for each product and each time using the equalization weight (S358). The smart set (s') calculated here is data different from the first smart set, and will be described below as a 'uniformized smart set'. When the homogenization smart set is calculated in this way, the third calculator 225 calculates individual weights for each product and for each time point using it (s359). Therefore, it is possible to calculate and allocate future demand forecasts for each product. Through this, it is possible to know from which pattern the future demand forecast was derived.

That is, in the process of FIG. 4, a smart set is calculated by adding past demand data generated at each point in time for each product based on past demand data generated during a predetermined period of time for one or more products, and inputting the calculated smart set information into the neural network , According to learning of the neural network, weights are calculated for each product at each time, the weights are equalized for each time, a smart set and a uniformed weight are calculated for each product to calculate a uniformed smart set, and individual weights for each product at each time point using the uniformed smart set The step of calculating and generating a future forecast value for each product is sequentially performed to generate a future demand forecast value and provide an explanation for the generated result.

The process of FIG. 4 is numerically expressed with reference to the table shown in FIG. 5 as follows. As shown in FIG. 5, assuming that demand data for 'Product 1' to 'Product 5' occurs intermittently for each time point '1' to '12', data showing such an intermittent pattern is converted into time series data so that there are no missing values. Transform to create smart set(s). Then, the time weight for each time point is calculated by the neural network model, and then the equalization weight is obtained using the softmax function. The sum of all equalization weight values is '1'.

Then, a uniform smart set (s') is obtained through the uniformization weight and smart set information, and then, through this, individual weights for each product at each point in time are obtained. Then, individual weights can be obtained for each product. This value becomes the future distribution value for each product. And through this, we can find the future prediction value by summing it. In FIG. 5, the future prediction value is '22.78081', and this value becomes the sum of future distribution values for each product or the sum of the uniformed smart set (s').

As such, the present invention generates demand forecasts for future time points regardless of time series data with steady patterns or intermittent data showing intermittent demand patterns. It can be seen that by providing forecast values for each product of , it is possible to provide interpretation data on which future forecast values were most affected at some point in the past.

Although the above has been described with reference to the illustrated embodiments of the present invention, these are only examples, and those skilled in the art to which the present invention belongs can variously It will be apparent that other embodiments that are variations, modifications and equivalents are possible. Therefore, the true technical protection scope of the present invention should be determined by the technical spirit of the appended claims.

10: Future demand forecast data generator
100: data preparation engine unit
110: account data generation unit
120: pre-processing unit
130: input information providing unit
200: prediction/analysis engine unit
221: smart set calculator
222: first calculation unit
223: function part
224: second calculator
225: third calculation unit
210: model fitting part
220: neural network model providing unit
230: result providing unit
240: correction unit

Claims (9)

a data preparation engine unit that prepares data for forecasting future demand using past demand data;
Includes a forecasting/analysis engine unit that predicts and analyzes future demand based on prepared data;
When the past demand data shows an intermittent pattern, the data preparation engine unit creates a smart set converted into time series data by adding past demand data generated at each point in time for each product,
The prediction/analysis engine unit decomposes future forecast data for intermittent demand into individual future forecast data using a neural network model and provides them,
The prediction/analysis engine unit includes a neural network model providing unit,
The neural network model providing unit,
a first calculation unit that calculates a time-specific weight for each product with respect to the generated smart set;
a function unit that equalizes the weights;
a second calculation unit calculating a homogenized smart set using the smart set and a homogenized weight;
A future demand forecasting data generating device comprising a third calculation unit that calculates individual weights for each product using the homogenized smart set, and providing a future demand forecast value for each product through the individual weights for each product. According to claim 1,
The data preparation engine unit,
an account data generation unit that sets a range of databases to be used for demand forecasting and creates accounts according to aggregation criteria;
Future demand forecasting data generating device further comprising a pre-processing unit for pre-processing the account. According to claim 1,
The neural network model is a neural network, characterized in that it is based on a recurrent neural network (RNN) using an attention mechanism. delete In the future demand forecast data generating device for generating demand forecast data at a future point using past demand data,
The future demand forecast data generating device,
converting intermittent demand pattern data into time series data;
learning the time series data with a neural network;
providing individual future demand forecast data for products according to learning results; and
Using the individual future demand forecast data to provide an analysis report on what pattern it originated from in the past,
The step of learning the neural network,
Calculating a smart set by adding past demand data for each product at each point in time;
inputting the calculated smart set information to the neural network;
Calculating a weight for each product according to learning of the neural network;
equalizing time-specific weights using a predetermined function;
calculating a homogenized smart set by calculating a smart set and a homogenization weight for each product;
A method for generating future demand forecast data comprising the step of calculating individual weights for each product at each point in time using the homogenized smart set and generating a future forecast value for each product. delete delete According to claim 5,
The step of learning the neural network,
A method for generating future demand forecast data learning using an Attention Mechanism-based Recurrent Neural Network (RNN) neural network. According to claim 5,
The function is a method for generating future demand forecast data, characterized in that the softmax function is applied.

Images