TFT-LCD產業階層式先進規劃排程---子計畫二：TFT-LCD產業之需求規劃

行政院國家科學委員會專題研究計畫 成果報告

TFT-LCD 產業階層式先進規劃排程--子計畫二：TFT-LCD 產 業之需求規劃(第 3 年)

研究成果報告(完整版)

計 畫 類 別 ： 整合型

計 畫 編 號 ： NSC 95-2221-E-011-137-MY3

執 行 期 間 ： 97 年 08 月 01 日至 98 年 07 月 31 日 執 行 單 位 ： 國立臺灣科技大學工業管理系

計 畫 主 持 人 ： 王福琨

計畫參與人員： 碩士班研究生-兼任助理人員：盧威利 碩士班研究生-兼任助理人員：曾治瑋 博士班研究生-兼任助理人員：鄭永福

報 告 附 件 ： 出席國際會議研究心得報告及發表論文

處 理 方 式 ： 本計畫涉及專利或其他智慧財產權，2 年後可公開查詢

中 華 民 國 98 年 10 月 09 日

行政院國家科學委員會補助專題研究計畫期末報告

產業階層式先進規劃排程‐子計畫二：TFT‐LCD 產業之需求規劃  Demand Planning in TFT‐LCD Industry

計畫類別：整合型計畫

計畫編號：NSC‐95‐2213‐E011‐137‐MY3

執行期限：95 年 8 月 1 日至 98 年 7 月 31 日

主持人：王福琨 參與人員：鄭永福、盧威立、曾治瑋 台灣科技大學工業管理系

Abstract

Improving the accuracy of demand forecasting has become a primary concern for a thin-film transistor liquid crystal display manufacturer. To address this concern, we develop a demand forecasting methodology that combines market and shipment forecasts. We investigate the weights assigned to the combination of forecasts using three linear methods (the minimum values of the forecast error, the adaptive weights and the regression analysis), as well as two nonlinear methods (fuzzy neural network and adaptive network based fuzzy inference system). A real data set from a panel manufacturer in Taiwan is used to demonstrate the application of the proposed methodology. The results show that the adaptive network based fuzzy inference system method outperforms other four methods. Also, we find that the mean absolute percent error (MAPE) of forecasting accuracy using the adaptive network based fuzzy inference system method can be improved effectively.

1. Introduction

Demand forecasting plays an important role for many global manufacturing enterprises.

Overestimating the demand will result in excess inventory, while underestimating will lead to lost sales. The accurate forecasting of future demand is a key when planning activities. For example, in short term production planning, a very detailed demand forecast of the size and location is required.

In either plant designing or budgeting, a precise forecast of the total market will be needed. To supply users (each representing different organizational functions and management echelons) with decision support information, the reliance for family based forecasting is stressed.

The inspiration for this study came from a thin-film transistor liquid crystal display (TFT-LCD) panel manufacturer in Taiwan.

Originally, the demand forecasting method was only based on a company’s shipment data, derived from a simple exponential smoothing model. In the past, the forecasting accuracy for all product levels was not sufficient in terms of the mean absolute percent error (MAPE), seeing as the MAPE for products at all levels was between 20%~30%. Such misrepresentations will result in excess inventory and lost sales.

Clearly, ways in which the accuracy of demand forecasting can be improved must be addressed.

In this study, we proposed a combined forecast using the fuzzy neural networks for a panel manufacturer. This article is divided into the following sections. The related literature of the research problem and the combined forecasting study are reviewed in Section 2; this is followed by the proposed methodology for demand forecasting. In Section 4, we analyze the proposed methodology with a real data set from a TFT-LCD panel manufacturer in Taiwan.

The conclusions, along with further research issues are made in the final section.

2. Demand forecasting in a panel manufacturer

A TFT-LCD manufacturing process consists of three basic process stages: array, cell, and module. The array process and processing steps are simpler versions of the semiconductor industry. The cell process consists of many steps that assemble TFT-array substrate, color filter (CF) substrate and polarizer. The module process is the last stage that determines the final product to be fit to delivery to the customer.

At the end, the stages are assembled together with all the necessary parts, including the drive IC (integrated circuit), PCB (printed circuit board), backlight unit and frame. The planning system of each process has different goals. For example, the array and cell processes are classified as capacity-oriented productions that emphasize the high utilization of machines; they also reduce the chance of capacity loss. This is so because such processes require expensive equipment.

However, the module process involves a material-oriented production environment that depends on the availability of key components including drive IC, PCB, backlight unit, frame, etc. Improving the accuracy of demand forecasting has become a primary concern for thin-film transistor liquid crystal display (TFT-LCD) manufacturers.

Lin et al. (2006) proposed a methodology for the supply and demand in the TFT-LCD market. A heuristic approach is used to forecast the future supply, while a transfer function model is used to forecast the future demand. The analysis of the supply and demand methodology shows that it could predict whether or not there appears to be a shortage in the market of 2004. Lo et al.

(2008) investigated the LCD monitor market using a hierarchical forecasting approach.

The results show that the best forecast results can be obtained by using the middle-out forecasting approach.

Individual forecasts can produced by forecasting models such as autoregressive integrated moving average (ARIMA), decomposition, transfer function, and the

Bass diffusion model with price variable.

They can also be produced by growth curve models such as Probit curve, Gompertz curve and the Logistic curve.

The Transfer function is also called a multiple input Box-Jenkins model (Box et al., 1994), which is a time series modeling process that describes a single dependent series as a function of its own past values and the values of one or more independent input series. As with univariate modeling, the purpose of multiple input modeling is to find the model that accounts for the predictable portion of the dependent series. Such a model can then be used for both forecasting and control. The Bass diffusion model can describe the empirical adoption curve quite well for many new products and technological innovations (Bass, 1969). The Bass diffusion model implies exponential growth of initial purchases to a peak and then exponential decay. This model provides good predictions of the timing and magnitude of the sales peak for the products to which it is applied. Bass et al. (1994) proposed a generalized Bass model, which includes decision variables such as price and advertising. The generalized model can be reduced to the Bass model in special cases; it also explains why the Bass model works so well without decision variables. Bass (1969) used the ordinary least squares (OLS) method to estimate the parameters.

Schmittlein and Mahajan (1982) used the maximum likelihood estimation (MLE) method to improve the estimation. Srinivsan and Mason (1986) used a nonlinear least square estimation (NLS) to obtain the valid error estimates. Growth curve models have been performed to forecast many different markets with successful results (Meade, 1984). The Gompertz and Logistic curve are widely used curves to forecast products.

Meade and Islam (1995) proposed the simple logistic for product forecasting, while the Gompertz models are shown to significantly outperform more complex models. The Probit model is another common curve utilized to forecast product demand (Daganzo, 1979). For these reasons, we

adopted all three growth curves to investigate the suitable forecasting model in this study.

The demand forecasting for a panel manufacturer consists of two stages. The descriptions of these two stages are as follows:

Stage 1 - To determine the suitable models for global market and company shipment forecasts, we use several forecasting methods such as ARIMA, Decomposition, Transfer function, Bass diffusion model with price variable and Growth curve models.

According to the objective of this study, we evaluate the best fitted model based on R-square value, Akaike information criterion (AIC) and MAPE.

Stage 2 - Forecast accuracy can be substantially improved through combining various forecast values into a composite forecast. The purpose of this stage is to combine forecasts with the linear method obtained from global and company shipment forecasts. Generally, a combined forecast will have a smaller forecast error. Several techniques such as the minimum values of the forecast error, the adaptive weights and the regression analysis can be used to obtain the weights. In this study, the weights can be determined to minimize the forecast error.

Bates and Granger (1969) developed and tested a number of techniques for combining point forecasts. They suggested a method that assumes the individual forecasts are consistent over time, a method that also minimizes the variance of the forecast errors over the covered time period. The weight assigned to the first forecast model, k, is given by

2 2 1

2 2 1

2 2 1

2

2

ρσ σ σ

σ

σ ρσ σ

− +

= −

k (4)

where =the variance of errors for the first model, =the variance of errors for the second model, and

2

σ

σ

ρ

suggested the best weighting scheme as one that involves allowing the weights to change from period to period. An adaptive set of weights are given by

∑−

− =

= ^T +

v T

t t t

v t

T e e

e

2 2 2 1

2 , 2

α

where

α

=

v the selective periods, and T =the total periods. Nelson (1984) suggested the regression method to combine forecasts which is given by

( )1 ₂ ( )2

1

* bF b F

F = +

where F^* = the combining forecast,

1 =

b the weight assigned to the first forecast model, and b₂ =he weight assigned to the second forecast model. Furthermore, Clemen (1989) provided a review and annotated bibliography of combining forecasts. A combining forecast has been shown to improve forecasting accuracy. With respect to the forecast accuracy of all methods, we adopted a mean absolute percentage error (MAPE). This metric was chosen because of its wide adoption ability to compare the performance of all methods.

3. Methodology

Besides the linear methods, Granger and Terasvirta (1993) suggested the nonlinear method may improve the performance of combining forecasting. Artificial neural network (ANN) and fuzzy neural network (FNN) are two common nonlinear techniques for combining forecasting in recent years.

They are trained on real data from a time series data and produce the forecasting based on others data and parameters. Donaldson and Kamstra (1996) combined the individual forecast by ANN. Fiordaliso (1998) proposed a FNN, which combined the concepts of neural network and fuzzy logic technique.

Palit and Popovic (2000) proposed a simulated hybrid model based on the isolated uses of neural networks and fuzzy logic, as

well as neuron-fuzzy systems. Dong (2002) used the adaptive-network-based fuzzy inference system (ANFIS) concept to combine several individual forecasts. Caydas et al. (2009) used the ANFIS model for the prediction of the white layer thickness (WLT) and the average surface roughness achieved as a function of the process parameters. The ANFIS, first proposed by Jang (1993), combined the benefits of ANN and fuzzy inference systems. Fuzzy inference system is the use of either prior experiences or knowledge into a set of constraints to obtain the optimize solution. The ANN structure can capture quite obvious patterns. ANFIS could adapt the parameters of the membership functions quickly and optimize them depending on the data being input.

Here, we investigated the nonlinear combined forecast which can be obtained by two individual forecasts and three linear combined forecasts obtained by ANFIS. In our proposed method, we assumed that the FIS have five inputs (the first forecasting model at period t), (the second forecasting model at period t), (the Bates and Granger’s combining model at period t), (the Bessler and Brandt’s combining model at period t) and (the linear composite model at period t); it also has output z (the best-fitted forecasts produced by ANFIS). The first-order Sugeno fuzzy model has become a common practice on ANFIS implements in the past.

Thus, we used the same model. The process shows as follows:

F_1,t

F_2,t

F_3,t

F_4,t

F_5,t

Layer 1: In this layer, each node is called an input linguistic node and corresponds to one input linguistic variable. The nodes transmit input forecasts to the next layer directly.

Each node function can be modeled by fuzzy membership function. Here, the generalized bell membership function and Gaussian membership function are used.

Layer 2: Each node in this layer calculates the firing strength of a rule via multiplication.

Layer 3: The ith node in this layer calculates

the ratio of the ith rule’s firing strength to the sum of all the rules’ firing strengths. The result would be the normalized firing strengths. For convenience, the output of this layer will be called the ‘normalized firing strengths’.

Layer 4: In this layer, each node i in this layer is a square node with a node function.

Parameters in this layer will be referred to as consequent parameters by node function.

Layer 5: The single node in this layer computes the final combining forecast as the summation of all incoming forecasts. All of the ANFIS functions were carried out in the mathematical software package MATLAB.

4. Case study

In general, a large TFT-LCD market can be divided in terms of application level (monitor, notebook, LCD-TV and industrial), size level (15-inches, 17-inches and 19-inches etc. for monitor product), resolution level, and product model. Because of the confidentiality entitled to the companies, we only investigated the size level of the LCD monitor market. The data set in Table 1 is recorded from the first quarter of 2000 to the fourth quarter of 2007. Also, Figures 3-5 show the trends for all different inches. The last two quarters are used as the holdout periods to test the forecasting accuracy.

To demonstrate the application of the proposed methodology, we presented the analysis results using the rolling concept. R0, the first analysis, is the data set from 2000/Q1 to 2007/Q2. It was used to obtain the forecasting fitted model. The data in 2007/Q3 was used to test one-period ahead forecasting accuracy. The second analysis is called R1, and it is the data set from 2000/Q1 to 2007/Q3; R1 was used to obtain the fitted forecasting model. The data in 2007/Q4 was used to test the forecasting accuracy of the period ahead.

According to the data from the company’s shipment and several forecasting models in Section 3, we found that the best-fit model for 15-inches was the Decomposition model; for the 17-inches and 19-inches, it was the Transfer function; for

the 20-inch, it was the Gompertz curve (see in Table 4). The R-square values for the four models are 97.31%, 90.71%, 90.66% and 91.96%, respectively. The MAPE values for the four models for the historical periods are 6.70%, 11.71%, 16.72% and 15.51%, respectively. The p-value of Kolmogorov-Smirnov test for normality of the error terms was used to check the model assumptions. The diagnostic checks for all fitted models validate the assumptions (see in Table 4).

To improve forecasting accuracy, the market and shipment forecasts were combined into a composite forecast. The results using five different combining methods in Sections 3-4 are shown in Table 5. Here, Method-1 is based on equation (2); Method-2 is based on equation (5); Method-3 is based on equation (6); Method-4 is based on a fuzzy neural network; and Method-5 is based on an adaptive network based fuzzy inference system. The comparison results show that the ANFIS method outperforms other four methods. The MAPE values for the four sizes for the historical periods using the ANFIS method are 3.29%, 0.21%, 0.12% and 0.02%, respectively. In addition, we can obtain the MAPE values of all four models for the historical periods using market forecasts from global market forecasts multiple market shares forecasts (see in Table 5). The results show that combining the forecasts is more effective than individual forecasts such as market forecasts and company’s shipment forecasts.

5. Conclusions

In the past, the forecasting accuracy from a specific (TFT-LCD) panel manufacturer in Taiwan was not adequate in terms of the mean absolute percent error (MAPE), since the percent error of the products at all levels of the company were between 20%~30%. This will result in excess inventory and lost sales. Thus, improving the accurate prediction of demand forecasting is an issue worth considering. The demand forecasting methodology proposed in this

study consists of two stages. First, we obtain the best fitted models for global market forecasts and company’s shipment forecasts from forecasting models such as ARIMA, Decomposition, Transfer function, and Bass diffusion model with price variable; Growth curve models such Probit curve, Gompertz curve and Logistic curve were also used. Then, we obtained the market forecasts using several forecasting methods such as Naive, growth curve model, moving average, exponential smoothing and decomposition. Second, we combined the market and shipment forecasts into a composite forecast so as to improve forecasting accuracy. Using five different combining methods, the comparison results show that the ANFIS method outperforms other four methods, including the minimum values of the forecast error, the adaptive weights, the regression analysis and the FNN method.

Also, the results show that the MAPE of forecasting accuracy can be improved effectively from 20%~30% to less than 5%.

Further research issue can be extended to the entire TFT-LCD market. In addition, consideration of the collaboration forecasting among the supplier, manufacturer, distributor, and retailer could result in more benefits in this TFT-LCD supply chain.

References

Bass, F. M. (1969). A new product growth model for consumer durables.

Management Science, 15, 215-227.

Bass, F. M., Krishnan, T. V. & Jain, D.

(1994). Why the Bass model fits without decision variables. Marketing Science, 13, 203-223.

Bates, J. M. & Granger, C. W. (1969). The combination of forecasts. Operational Research Quarterly, 20, 451-468.

Bessler, D. A. & Brandt, J. A. (1981).

Composite forecasting: an application with U.S. hog price. American Journal of Agricultural Economics, 63, 135-140.

Box, G. E. P., Jenkins, G. W. & Reinsel, G.

C. (1994). Time Series Analysis Forecasting and Control. Englewood Cliffs, NJ: Prentice-Hall.

Caydas, U., Hascalik, A. & Ekici, S. (2009).

An adaptive neuro-fuzzy inference system (ANFIS) model for wire-EDM. Expert Systems with Applications, 36, 6135-6139.

Clemen, R. T. (1989). Combining forecasts:

a review and annotated. International Journal of Forecasting, 5, 559-583.

Daganzo, C. (1979). Multinomial Probit, the Theory and its Application to Demand Forecasting. New York: Academic Press.

Donaldson, R. G. & Kamstra, M. (1996).

Forecast combining with neural networks.

Journal of Forecasting, 15, 49-61.

Dong, J. R. (2002). A nonlinear combining forecast method based on fuzzy neural network,” Proceedings of the First International Conference on Machine Learning and Cybernetics, Beijing, pp.

2160-2164.

Fiordaliso, A. (1998). A nonlinear forecasts combination method based on

Takagi–Sugeno fuzzy systems.

International Journal of Forecasting, 14, 367-379.

Granger, C. W. J. & Terasvirta, T. (1993).

Modelling Nonlinear Economic Relationships. New York: Oxford University Press.

Jang, J. S. R. (1993). ANFIS:

adaptive-network-based fuzzy inference system. IEEE Transactions on Systems, Man, Cybernetics, 23, 665-685.

Lin, J. T., Wang, F. K., Lo, S. L., Hsu, W. T.

& Wang, Y. T. (2006). Analysis of the supply and demand in the TFT-LCD market. Technological Forecasting and Social Change, 73, 422-435.

Lo, S. L., Wang, F. K. and Lin, J. T. (2008).

Forecasting for the LCD monitor market.

Journal of Forecasting, 27, 341-356.

Meade, N. (1984). The use of growth curves in forecasting market development – a review and appraisal. Journal of Forecasting, 3, 429-451.

Meade, N. & Islam, T. (1995). Forecasting with growth curves: an empirical comparison. International Journal of

Forecasting, 11, 199-215.

Nelson, C. R. (1984). A benchmark for the accuracy of econometric forecasts of GNP.

Business Economics, 19, 52-58.

Palit, A. K. & Popovic, D. (2000). Nonlinear combination of forecasts using artificial neural networks, fuzzy logic and neuro-fuzzy approach. The Ninth IEEE International Conference of Fuzzy Systems, pp. 566-571.

Robinson, B. & Lakhani, C. (1975). Dynamic price models for new product planning.

Management Science, 10, 1113-1122.

Schmittlein, D. C. & Mahajan, V. (1982).

Maximum likelihood estimation for an innovation diffusion model of new product acceptance. Marketing Science, 1, 57-78.

Srinivasan, V. & Mason, C. H. (1986).

Nonlinear least squares estimation of new product diffusion models. Marketing Science, 5, 169-178.

Self Evaluation

During these three years, we have been finished three papers. One paper has been accepted by the journal “Expert Systems with Applications (SCI)”. Two papers have been submitted to SCI journals for reviewing.

一、參加會議經過

The 29th International Symposium on Forecasting (ISF) 學會的成員主要是全 球在預測研究領域的學者。本屆學術研討會地點於香港舉行，時間為 6 月 21 日 至 6 月 24 日。本人原本於會議中發表論文”Using fuzzy neural networks for  combined forecasts”。但由於臨時生病故無法參加。故此會議僅繳交報名費，會 後有一本電子版的論文集。補助出國會用僅包括報名費。

報告人：台灣科技大學工業管理系王福琨教授