Automatic Window Selection for Nonstationary Time Series Forecasting
비정상 시계열 예측을 위한 자동 윈도우 선택 모델
- 주제(키워드) Nonstationary time-series forecasting , Convolution long short-term memory , Self-attention , Automatic time window
- 발행기관 고려대학교 대학원
- 지도교수 백준걸
- 발행년도 2022
- 학위수여년월 2022. 2
- 학위구분 석사
- 학과 대학원 산업경영공학과
- 원문페이지 44 p
- UCI I804:11009-000000257386
- DOI 10.23186/korea.000000257386.11009.0001252
- 본문언어 영어
초록/요약
Time-series data are generated and obtained from various industries. Therefore, they exhibit various patterns, which significantly affect prediction performance. Most time-series data obtained in industrial fields exhibit nonstationary characteristics that are difficult to predict. Nonstationary time-series data exhibit uncertain pattern characteristics. Existing methods for time-series prediction use static time window sizes, regardless of the pattern characteristics, for convenience. The size of the static time window does not reflect all temporal information in the historical data. Therefore, it is not suitable for nonstationary data with uncertain pattern characteristics. In this paper, propose a multiple window map and a new methodology applied to self-attention and Convolution Long Short-Term Memory (C-LSTM). The proposed method uses a multiple window map to represent windows of all sizes in the historical data. In addition, through self-attention, an appropriate time window size is automatically selected to predict the next time step. This guarantees the reliability and accuracy of the prediction results of nonstationary time-series data. In addition, by enabling end-to-end learning using data transformation and C-LSTM, the learning efficiency is increased. The performance superiority and high learning efficiency of the proposed method are verified experimentally on various nonstationary real data. In addition, the results of automatic time window size reflection based on data patterns are confirmed through a qualitative analysis of the attention distribution map.
more목차
1. Introduction 1
2. Proposed method 7
2.1 Block 1: Data transformation and feature extraction 8
2.2 Block 2: Attention weights calculation 11
2.3 Block 3: LSTM for time series predictions 14
3. Experiments 14
3.1 Dataset 14
3.2 Evaluation metrics 16
3.2 One step prediction result 17
3.2 Training efficiency measurement results 21
3.2 Multi step prediction results 24
3.2 Qualitative analysis of attention weights 26
4. Conclusion 28
References 29

