臺灣博碩士論文加值系統

間歇性需求數據通常稱為客戶需求數據或銷售數據。數據集會將有需求之資料記錄為一非零值；若無需求，則數據集記錄為零值。然而，一般常見的需求並非是連續性的，而是間歇性的資料，此特性使得間歇性數據難以進行預測。常見用於預測間歇性需求數據的標準方法包含Croston、單指數平滑（SES）等等，而 Croston和SES此兩種方法常用於靜態之預測。
此研究利用深度學習中的遞歸神經網絡 (Recurrent Neural Network (RNN))、門基循環單元 (Gated Recurrent Unit (GRU)) 和長期短期記憶 (Long Short Term Memory (LSTM)) 來預測間歇性數據。此模擬研究將重複生成50次6個不同設計參數之數據集，此外，實證研究上，也以Kaggle網站上之數據做為資料集。本研究以RNN、GRU及LSTM預測間歇性需求數據的結果，並與Croston和SES作為基準方法進行比較。評估結果的方法則是以平均絕對誤差 (Mean Absolute Error (MAE)) 和均方根比例誤差(Root Mean Squared Scaled Error (RMSE)) 作為評斷指標。
在模擬研究中，大多數遞歸神經網絡方法在MAE評分中表現良好。 對於經驗研究，遞歸神經網絡方法在所有數據集的MAE評分方面均優於傳統方法。然而，對於大多數模擬研究和一項實證研究，Croston的常規方法都適用於RMSSE分數。

Intermittent demand data is usually called customer demand data or sales data. The dataset will record a nonzero value if there is a demand. If there is no demand, the dataset records are zero values. The general problem is that demand is not always continuous but intermittent. This characteristic makes intermittent data difficult to use for prediction. Standard methods used to predict intermittent demand data include Croston, single exponential smoothing (SES), and others. The Croston and SES typically produce static forecasts. This study utilized deep learning methods recurrent neural network (RNN), gated recurrent units (GRU), and long short-term memory (LSTM) to predict intermittent data. The simulation study was carried out by generating datasets with 6 different design parameters and with 50 repetitions. Besides, the empirical study used data from the Kaggle website. This study measured the performance of predicting intermittent demand data by RNN, GRU, and LSTM, comparison to Croston and SES as the benchmark methods. The performance measurements included the evaluation of mean absolute error (MAE) and root mean squared scaled error (RMSSE). In simulation studies, most recurrent neural network methods can perform well in MAE scores. For the empirical study, recurrent neural network methods outperform conventional methods in MAE scores for all datasets. Yet, the convention method of Croston works in RMSSE scores for most simulation studies and one empirical study.

摘要 . . .i
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .ii
Acknowledgement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .iii
Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1 Conventional Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.1 Croston Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.2 Simple Exponential Smoohting (SES) . . . . . . . . . . . . . . . . . . 5
2.2 Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.1 Recurrent Neural Network (RNN) . . . . . . . . . . . . . . . . . . . . 8
2.2.2 Gated Recurrent Unit (GRU) . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.3 Long Short­Term Memory (LSTM) . . . . . . . . . . . . . . . . . . . 10
2.3 Bayesian Optimization by Keras­Tuner . . . . . . . . . . . . . . . . . . . . . 13
2.4 Forecasting Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3 Materials and Implement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.1 Intermittent Demand Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.1.1 Simulation Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.1.2 Empirical Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.2 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2.1 Predict Second half . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2.2 Evaluation Criterions . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4 Result and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.1 Simulation Study Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.2 Empirical Study Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

[1] J. D. Croston, “Forecasting and stock control for intermittent demands,” Journal of the
Operational Research Society, vol. 23, no. 3, pp. 289–303, 1972. [Online]. Available:
https://doi.org/10.1057/jors.1972.50
[2] R. D. Snyder, J. K. Ord, and A. Beaumont, “Forecasting the intermittent demand
for slow­moving inventories: A modelling approach,” International Journal of
Forecasting, vol. 28, no. 2, pp. 485 – 496, 2012. [Online]. Available: http:
//www.sciencedirect.com/science/article/pii/S0169207011000781
[3] S. Viswanathan, H. Widiarta, and R. Piplani, “Forecasting aggregate time series with inter­
mittent subaggregate components: Top­down versus bottom­up forecasting,” Ima Journal
of Management Mathematics ­ IMA J MANAG MATH, vol. 19, pp. 275–287, 03 2007.
[4] N. Kourentzes, “Intermittent demand forecasts with neural networks,” International Jour­
nal of Production Economics, vol. 143, pp. 198–206, 05 2013.
[5] Y. Hong, J. Zhou, and M. Lanham, “Forecasting intermittent demand patterns with time
series and machine learning methodologies,” 2018.
[6] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, Learning Representations by Back­
Propagating Errors. Cambridge, MA, USA: MIT Press, 1988, p. 696–699.
[7] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. The MIT Press, 2016.
[8] K. Cho, B. van Merriënboer, C. Gulcehre, F. Bougares, H. Schwenk, and Y. Bengio,
“Learning phrase representations using rnn encoder­decoder for statistical machine trans­
lation,” 06 2014.
[9] A. Syntetos and J. Boylan, “On the bias of intermittent demand estimates,” International
Journal of Production Economics, vol. 71, pp. 457–466, 05 2001.
[10] E. Ostertagova and O. Ostertag, “The simple exponential smoothing model,” 09 2011.

[11] R. Hyndman and G. Athanasopoulos, Forecasting: Principles and Practice, 2nd ed. Aus­
tralia: OTexts, 2018.
[12] P. Zhang, E. Patuwo, and M. Hu, “Forecasting with artificial neural networks: The state
of the art,” International Journal of Forecasting, vol. 14, pp. 35–62, 03 1998.
[13] E. Brochu, V. Cora, and N. Freitas, “A tutorial on bayesian optimization of expensive
cost functions, with application to active user modeling and hierarchical reinforcement
learning,” CoRR, vol. abs/1012.2599, 12 2010.
[14] M. Syntetos, J. Boylan, and J. Croston, “On the categorization of demand patterns,” Jour­
nal of the Operational Research Society, vol. 56, 05 2005.
[15] Hendra, H. Napitupulu, H. Nasution, and J. Hidayati, “Portfolio effect on end user spare
parts based on demand patterns,” IOP Conference Series: Materials Science and Engi­
neering, vol. 505, p. 012075, 07 2019.
[16] F. Petropoulos, S. Makridakis, V. Assimakopoulos, and K. Nikolopoulos, “‘horses for
courses ’in demand forecasting,” European Journal of Operational Research, vol. 237, p.
152–163, 08 2014.
[17] R. J. Hyndman and A. B. Koehler, “Another look at measures of forecast accuracy,”
International Journal of Forecasting, vol. 22, no. 4, pp. 679 – 688, 2006. [Online].
Available: http://www.sciencedirect.com/science/article/pii/S0169207006000239