TWI771615B - Product order prediction method - Google Patents

Abstract

A product order prediction method adapts to a production planning system. The method comprises obtaining a reference data related to a next reference product order and a current actual product order, performing an algorithm based on neural network model according to the reference data and the current actual product order to generate a feature vector, and performing another algorithm based on neural network model according to the feature vector to output a next predicted product order to the production planning system, for the production planning system to plan an execution of the production line of the product according to the next predicted product order.

Description

Order forecast method

The present invention relates to a method for order prediction, in particular to a method for predicting orders using a model trained by a neural network.

For electronics manufacturers, the ideal situation is to maintain inventory below a certain amount. Excessive inventory and low inventory turnover will take a lot of money and put pressure on cash flow. In addition, high inventory also adds additional The time and labor costs are used for warehousing expenses, picking products and counting.

Most of the current methods of order forecasting rely on experience, previous orders and inventory data to make judgments. However, humans cannot accurately grasp overly complex and high-dimensional data. Even if simple regression or time series analysis is used to book orders, they can only make predictions based on the previous situation, and cannot be analyzed from a long-term perspective to ensure accurate predictions. accuracy. Therefore, a high-accuracy order prediction method is currently in great demand.

In view of this, the present invention proposes an order prediction method based on a neural network-like model, thereby solving the problem of low accuracy of order prediction in the existing method.

An order forecasting method described according to an embodiment of the present invention is applicable to a product production planning system. The method includes: obtaining reference information related to the next reference order of the product and the actual order of the product; executing an algorithm based on a neural network model according to the reference information and the actual order to generate a feature vector; and executing another algorithm based on the neural network model according to the feature vector to output the next predicted order to the product production planning system for the product production planning system to plan the operation of the production line to produce the product according to the next predicted order.

With the above structure, the order forecasting method disclosed in the present invention can comprehensively consider various types of information related to an order to make a more accurate forecast for the next order. For example: if it is predicted that a large number of orders will be received next week, it will actively increase production capacity this week in advance to meet next week's orders; if it is predicted that the decrease in the number of orders next week will lead to an increase in the inventory rate, reduce production capacity this week to Avoid overproduction. The invention helps to reduce inventory, and further reduces the time cost and labor cost of inventory and storage.

The above description of the present disclosure and the following description of the embodiments are used to demonstrate and explain the spirit and principle of the present invention, and provide further explanation of the scope of the patent application of the present invention.

The detailed features and advantages of the present invention are described in detail below in the embodiments, and the content is sufficient to enable any person skilled in the relevant art to understand the technical content of the present invention and implement it accordingly, and according to the content disclosed in this specification, the scope of the patent application and the drawings , any person skilled in the related art can easily understand the related objects and advantages of the present invention. The following examples further illustrate the viewpoints of the present invention in detail, but do not limit the scope of the present invention in any viewpoint.

The order prediction method proposed by the present invention is suitable for a product production planning system. The product production planning system is, for example, a cloud server, which can be used to control the operation of the product production line. The product production planning system may include a processor and a memory, and may further include a hard disk and a plurality of computing units capable of parallel processing. The processor of the product production planning system runs the order forecasting method proposed by the present invention. The product production planning system adaptively adjusts the operation parameters of the machines and tools on the production line according to the output result of the order forecasting method proposed by the present invention, thereby increasing the production cycle or reducing the production cycle to achieve the effect of controlling product output. In practice, the product production planning system is, for example, an enterprise resource planning system (Enterprise Resource Planning, ERP), a purchase, sales, inventory (Purchase, Sales, Inventory, PSI) management system.

Please refer to FIG. 1 , which is a flowchart illustrating an order prediction method according to an embodiment of the present invention.

Please refer to step S1 to obtain the actual order and reference information of the product.

To be more specific, this actual order is, for example, the value recorded in the Sales Order (SO). The reference information is associated with the next reference order for the product. Reference information includes one or more reference values. If the reference information is a reference value, the reference value is, for example, the next reference order. If the reference information is a plurality of reference values, the reference values respectively correspond to a plurality of reference orders, one of these reference orders is the next reference order, and these reference orders are consecutive in time according to the first cycle. The first cycle is, for example, in a “week” unit, and the reference information is, for example, the reference order for the next week and the reference orders for consecutive N-1 weeks thereafter (that is, the reference information is the reference order for consecutive N weeks) . The reference information may further include the target inventory rate for N consecutive weeks, or the forecast price of raw materials for N consecutive periods, the weather for N consecutive weeks, or any combination of the above enumerations. The types and numbers of information that can be used as reference are not particularly limited.

Please refer to step S2, and according to the reference information and the actual order of this time, an algorithm based on a neural network model is executed to generate a feature vector. In an embodiment of the present invention, the algorithm based on the neural network model is, for example, a Long Short-Term Memory (LSTM) network including an attention mechanism (Attention mechanism), and its input parameters are A multi-dimensional vector (such as N+1 dimension), which consists of reference information (such as reference orders for N consecutive weeks) and this actual order (such as the cumulative order volume from last Sunday to this Saturday). The neural network-like model using LSTM takes the aforementioned multi-dimensional (eg, 30-dimensional) vector as the input layer parameter, and generates a multi-dimensional feature vector (eg, 64-dimensional).

In practice, the LSTM-like neural network model is pre-trained based on past reference information and past actual orders. For example, using information such as reference orders and actual orders from week 1 to week 20 to automatically adjust multiple weights in the neural network model, and use the trained model in the product production planning system and production line in week 21 Operations, forecasts and actual orders made from week 21, optionally fed back into the neural network-like model. It should also be mentioned that the reference information can be input to the neural network-like model using LSTM using different cycles. For example, the reference order uses the "week" unit as the input layer parameter, and the raw material forecast price uses the "month" unit as the input layer parameter. When updating the input layer parameters of the reference order every week, if it is not the time when the forecast price of raw materials is updated every month, you can optionally enter a 0 value or continue the previous input value instead.

Referring to step S3, another algorithm is executed according to the feature vector to output the next predicted order to the product production planning system, so that the product production planning system can control the operation of the production line to produce products according to the next predicted order. Said other algorithm is an algorithm based on a type of neural network model, or an algorithm based on a regression analysis. In practice, another algorithm based on the neural network-like model described above is a multi-layer perceptron (MLP). MLP takes the multi-dimensional feature vector generated by LSTM as an input parameter, and converts the feature vector into a scalar by means of linear combination, which represents the next prediction order. Therefore, the production line control system can adaptively adjust the production progress of the subsequent production line according to the value of the next predicted order, thereby avoiding an increase in the inventory rate caused by additional production. In another embodiment of the present invention, the MLP may also output a plurality of scalars representing consecutive forecast orders. For example, if the forecast period is 4, the forecast orders for four consecutive weeks will be output, which will be used as the reference index of the product production planning system for the next month. Various production machines on the production line can also determine their own parameter settings during operation according to the next forecasted order.

Please refer to FIG. 2 , which is a line chart of “order volume-time” of the next predicted order, the next reference order and the actual order obtained by applying the order forecasting method according to an embodiment of the present invention, of which the first 20 weeks are The training phase of the neural network-like model, so the values of the actual order, the predicted order and the reference order are the same. The 21st week is the online operation stage of the neural network model. It can be seen from FIG. 2 that the predicted orders obtained by applying the order prediction method according to an embodiment of the present invention are getting closer and closer to the actual orders after about the 50th week. The broken line indicates that the accuracy of the order prediction method according to an embodiment of the present invention can be further improved with the online operation time.

To sum up, the order prediction method proposed by the present invention uses LSTM and MLP as the weight area and the analysis area respectively, LSTM outputs multi-dimensional vector information as the input of MLP, and MLP outputs one-dimensional (or multi-dimensional) vector information as the lower part. forecast orders. The input data of LSTM includes this actual order and reference information, and the reference information can cover various types of forecast information, such as reference orders for six consecutive months, forecast stock rate, etc. The order prediction method of the present invention can automatically adjust the weight value in the weight area during the training process, and after the online operation, it can continue to feed back to the neural network-like model as new training data according to the actual operation results and errors. Therefore, the accuracy of prediction can be further improved. Overall, the order forecasting system disclosed in this case can comprehensively consider various types of information related to product production and make more accurate forecasts for the next order. For example: if it is predicted that a large number of orders will be received next week, it will actively increase production capacity this week in advance to meet next week's orders; if it is predicted that the decrease in the number of orders next week will lead to an increase in the inventory rate, reduce production capacity this week to Avoid overproduction. The invention also helps to reduce inventory, and further reduces the time cost and labor cost of inventory and warehousing.

Although the present invention is disclosed in the foregoing embodiments, it is not intended to limit the present invention. Changes and modifications made without departing from the spirit and scope of the present invention belong to the scope of patent protection of the present invention. For the protection scope defined by the present invention, please refer to the attached patent application scope.

S1~S3: Steps

FIG. 1 is a flowchart illustrating an order prediction method according to an embodiment of the present invention. Figure 2 is a line chart of the "order volume-time" of the next forecast order, the next reference order and the actual order this time.

S1~S3: Steps

Claims (5)

An order forecasting method, applicable to a product production planning system, the method comprising: obtaining reference information related to a next reference order of a product and a current actual order of the product; This actual order executes an algorithm based on a neural network model to generate a feature vector; and executes another algorithm according to the feature vector to output the next predicted order to the product production planning system for the production of the product The planning system controls the operation of a production line to produce the product according to the next forecast order; wherein the other algorithm further outputs another next forecast order, and the next forecast order and the other next forecast order are separated by a forecast period; wherein The next reference order has nothing to do with the current actual order, and the next reference order is related to information after the time point of the current actual order. The order prediction method according to claim 1, wherein the algorithm based on the neural network model is a long short-term memory network. The order forecasting method of claim 1, wherein the other algorithm is an algorithm based on another type of neural network model or an algorithm based on a regression analysis. The order prediction method of claim 3, wherein the algorithm based on the other type of neural network model is a multilayer perceptron. The order forecasting method according to claim 1, wherein the reference information includes a plurality of reference values, the reference values correspond to a plurality of reference orders, and the reference orders are consecutive in time according to a first cycle, and the reference values One of the orders is the next reference order.

Images