낙동강 유역의 SPI 예측을 위한 심층신경망 개발 및 검증 = Development and validation of a deep neural network for predicting SPI of Nakdong river basin

국문 초록 (Abstract)

본 연구에서, 낙동강 유역의 SPI6를 예측하기 위해 신경망 모형 중 하나인 다층 퍼셉트론 (MLP; multilayer perceptron)을 적용했다. 월별 표준강수지수, 기온, 기온 평년값, 강수량, 강수 평년 값, 강수일수, 그리고 여러가지 세계기후지수를 설명 변수로 사용했으며, 최적의 모형을 찾기 위해 두 가지의 활성함수, 네 가지의 은닉층별 뉴런 수, 다섯 가지 중도탈락 비율 (dropout rate), 다섯 가지 은닉층 수, 두 가지 손실 함수를 고려하였다. 모형 평가 기준으로 검증평균제곱오차 (validation mean square error)를 이용하여 가장 좋은 모형을 찾고자 하였다. SPI6를 예측함에 있어서, (L_1) 정규화를 적용한 모형이 적용하지 않은 모형보다 검증평균제곱오차 (validation mean square error)가 항상 더 낮았으며, 대부분의 지역에서 은닉층 수가 적을수록 검증평균제곱오차 (validation mean square error)가 더 낮은 경향을 보였다.

다국어 초록 (Multilingual Abstract)

In this study, we applied a kind of deep neural network, a multilayer perceptron (MLP), to predict SPI6 of Nakdong River Basin. The monthly standard precipitation indices, temperature, temperature normal value, precipitation, normal precipitation, precipitation date, and various global climate indexes are used as explanatory variables. To find the optimal model, we consider two active functions, four types of the neurons per a hidden layer, five types of dropout rates, five types of the number of hidden layers, and two loss functions. We use the validation mean square error for the evaluation. The validation mean square error was always lower in the model with L_1 normalization than in the model without the normalization. In most areas, the smaller the number of hidden layers, the lower the validation mean square error.

참고문헌 (Reference)

1 Mckee, T. B., "The relationship of drought frequency and duration to time scales" 179-184, 1993

2 Nair, V., "Rectified linear units improve restricted boltzmann machines" 27 : 807-814, 2010

3 The H2O.ai team, "R interface for h2o"

4 Bonaccorso, B., "Probabilistic forecasting of drought class transitions in Sicily (Italy) using Standardized Precipitation Index and North Atlantic Oscillation Index" 526 : 136-150, 2015

5 Minsky, M., "Perceptions. An introduction to computational geometry" The MIT Press 1988

6 Goodfellow, I., "Maxout networks" 28 : 1319-1327, 2013

7 Dozat, T., "Incorporating nesterov momentum into adam" 1 : 2013-2016, 2016

8 Ali, M., "Improving spi-derived drought forecasts incorporating synoptic-scale climate indices in multi-phase multivariate empirical mode decomposition model hybridized with simulated annealing and kernel ridge regression algorithms" 576 : 164-184, 2019

9 Wickham. H., L., "Ggplot2: Elegant graphics for data analysis" Springer-Verlag 2016

10 Wu, H., "Appropriate application of the standardized precipitation index in arid locations and dry seasons" 27 : 65-79, 2007

11 Duchi, J., "Adaptive subgradient methods for online learning and stochastic optimization" 12 : 2121-2159, 2011

12 Kingma, D. P., "Adam: A method for stochastic optimization"

13 Matthew D. Zeiler, "Adadelta: An adaptive learning rate method"

14 Nesterov, Y., "A method for unconstrained convex minimization problem with the rate of convergence o(1/k2)" 269 : 543-547, 1983