Over the last few decades, meteorological and climatological predictions using weather and climate models have been improving steadily. In contrast, sub-seasonal forecasts, which are between weather and climate models, have revealed a gap in the construction of seamless numerical forecasting systems. Climate data produced mainly by the APEC Climate Center (APCC) is seasonally forecasted monthly data with a lead-time of one to six months. It cannot be used directly in hydrological or other application fields because of its low spatial and temporal resolution. In the meteorological / climatic field, efforts are being made to increase the predictability of the weather and the availability climate data by extending seasonal forecast data to the sub-seasonal time scale, from two weeks to two months. As the result, the Subseasonal-to-Seasonal (S2S) project conducted jointly by the World Weather Research Program and the World Climate Research Program of the World Meteorological Organization was begun in 2013.

In this study, we evaluated the accuracy of S2S forecast data, which has increased in importance recently given the interest in seamless weather and climate forecast data, and evaluated its applicability to disaster areas, especially in river flood risk assessment. To evaluate the accuracy of the S2S forecast data, the Namgang dam basin was selected as the study area and the grid rainfall data corresponding to the eight rainfall stations in the area were extracted in 40 and 7 day intervals. The prediction accuracy of the S2S forecast data was evaluated by direct comparison with the rainfall data observed at the eight stations. Analysis using the RMSE and the coefficient of determination (R2) were performed to evaluate the prediction accuracy of the S2S rainfall forecast data. To test the possibility of improving accuracy of this data, a study was conducted using the S2S rainfall forecast data for the Korean peninsula to predict the precipitation at Namgang Dam Rainfall Observation Station using the machine learning technique known as multilayer perceptron.

In addition, time series data for fluctuations in river water elevation in the Nam River dam were predicted using the Long Short-Term model (LSTM) technique, which is effective for applying machine learning to time series data, to evaluate the river flood risk using S2S prediction data. To evaluate the feasibility of estimating fluctuations in river water elevation using S2S prediction data, three variables were estimated using the LSTM method: 1) the river water elevation using only water elevation; 2) river water elevation using only daily precipitation data; and 3) changes in river level using daily composite (river water and precipitation) data.

The result of the study shows that if the LSTM model is constructed using the S2S prediction data and river water elevation, the risk of flooding can be understood in terms of disaster preparedness. However, the R2 value was very low from the 2nd day due to limitations in the predictability of rainfall within the S2S forecast data itself. Therefore, it is necessary to develop methods to improve the accuracy of the S2S forecast data for extension to areas such as agriculture, water resources, and energy.

1. 서론

2. 연구자료 및 방법

2.1 계절내 예측자료

2.2 남강댐 유역 수문자료

2.3 인공신경망

2.3.1 다층 퍼셉트론

2.3.2 활성화 함수

2.3.3 손실 함수

2.3.4 오차역전파 알고리즘

2.3.5 LSTM

2.3.6 Keras

2.3.7 학습모델 구축

2.4 계절내 예측자료의 정확도 평가 방법

3. 연구내용

3.1 계절내 예측자료의 강우 예측능력 평가

3.2 다층 퍼셉트론을 이용한 계절내 예측정보의 예측정확도 향상 평가

3.3 홍수위험도 예측 활용성 평가

3.3.1 시간 단위 하천수위를 이용한 하천수위 예측

3.3.2 일 단위 하천수위를 이용한 하천수위 예측

3.3.3 일 단위 관측강수를 이용한 하천수위 예측

3.3.4 일 단위 계절내 예측자료를 이용한 하천수위 예측

3.3.5 일 단위 복합자료를 이용한 하천수위 예측

4. 결론

REFERENCES