상세정보
학술대회
Uncertainty-aware Deep Learning Forecast using Dropout-based Ensemble Method
- 발행일
- 201910
- 출처
- International Conference on Information and Communication Technology Convergence (ICTC) 2019, pp.1120-1125
- 협약과제
- 19PK1100, 전력 빅데이터를 활용한 신산업 BM 및 서비스 개발·검증, 이병탁
- 초록
- A probabilistic forecast technique provides useful information showing how confident a point forecast is. However, building a good probabilistic forecast technique that is easily applicable to existing deterministic forecast techniques is not easy because it may require a case-specific or complicated method, as we observed in existing studies. In this paper, we propose a probabilistic forecast technique that is generally applicable to existing deterministic forecast techniques. Focusing on deep learning, which has shown promising results in various areas including forecasting, we achieve a generality by using a dropout technique that is widely applied in deep learning architectures. We first show how the dropout technique is utilized to realize an ensemble method. Then, through experiments using the real-world data, we describe an uncertainty metric, which is related to the forecast accuracy, and show that a lower uncertainty metric value is likely to indicate a higher forecast accuracy.
- KSP 제안 키워드
- Deep Learning Architectures, Ensemble method, Forecast Accuracy, Point forecast, Probabilistic Forecast, Real-world data, deep learning(DL), uncertainty metric