In this study, the applicability of a deep-learning LSTM model, which is based on the relationship between input and output data, was tested for forecasting non-linear time series data such as river water level and discharge. To achieve the objectives, three applications was carried out: (1) forecas...
In this study, the applicability of a deep-learning LSTM model, which is based on the relationship between input and output data, was tested for forecasting non-linear time series data such as river water level and discharge. To achieve the objectives, three applications was carried out: (1) forecasting of downstream water level in a natural river using only upstream water level data, (2) forecasting of tide-dominated river water level using tidal level and dam release data, (3) forecasting of monthly discharge using rainfall, temperature at the ASOS observation stations, and dam release data of the three upstream dams. The major findings are as following.
(1) As a result of forecasting the water level of the Okcheon station using the water level data of the three upstream water level stations, the forecasting capability was found that NSE values almost close to 1.0 in the case the where optimum model performance was exhibited.
(2) As a result of forecasting the water level of Jamsu Bridge, affected by sea water level, the forecast accuracy for the 1 hour lead time when an LSTM model exhibit optimum performance was found to be RMSE of 0.065 m and NSE of 0.99. However, when the lead time became longer, the forecasting accuracy error increased on average irrespective of the sequence length parameter.
(3) As a result of forecasting the monthly flow at the Han River Bridge, the case where monthly flow was included during the deep neural network learning showed better accuracy than that of the case where monthly flow was excluded. In addition, the case where only meteorological information was utilized in order to predict monthly discharge, the accuracy was shown to have been degraded due to over-fitting when the iteration number of learning increased. In particular, the forecasting model for the monthly flow of the current month (t) based on past time series(t-n ~ t) showed far superior to the model forecasting the flow of the following month (t+1).
In current hydrological forecast systems, a forecast usually carried out utilizing predictive hydrological data. However, such systems can not be trusted unless reliable parameters is not guaranteed for sufficient lead time. Therefore, if the purpose is to analyze the non-linear hydrological time series at a specific outlet that is naturally or artificially regulated, the LSTM model that remembers long-term information based on big data and reflects it on forecast after adjusting it is expected to be utilizable and applicable in various hydrological time series analysis.
In this study, the applicability of a deep-learning LSTM model, which is based on the relationship between input and output data, was tested for forecasting non-linear time series data such as river water level and discharge. To achieve the objectives, three applications was carried out: (1) forecasting of downstream water level in a natural river using only upstream water level data, (2) forecasting of tide-dominated river water level using tidal level and dam release data, (3) forecasting of monthly discharge using rainfall, temperature at the ASOS observation stations, and dam release data of the three upstream dams. The major findings are as following.
(1) As a result of forecasting the water level of the Okcheon station using the water level data of the three upstream water level stations, the forecasting capability was found that NSE values almost close to 1.0 in the case the where optimum model performance was exhibited.
(2) As a result of forecasting the water level of Jamsu Bridge, affected by sea water level, the forecast accuracy for the 1 hour lead time when an LSTM model exhibit optimum performance was found to be RMSE of 0.065 m and NSE of 0.99. However, when the lead time became longer, the forecasting accuracy error increased on average irrespective of the sequence length parameter.
(3) As a result of forecasting the monthly flow at the Han River Bridge, the case where monthly flow was included during the deep neural network learning showed better accuracy than that of the case where monthly flow was excluded. In addition, the case where only meteorological information was utilized in order to predict monthly discharge, the accuracy was shown to have been degraded due to over-fitting when the iteration number of learning increased. In particular, the forecasting model for the monthly flow of the current month (t) based on past time series(t-n ~ t) showed far superior to the model forecasting the flow of the following month (t+1).
In current hydrological forecast systems, a forecast usually carried out utilizing predictive hydrological data. However, such systems can not be trusted unless reliable parameters is not guaranteed for sufficient lead time. Therefore, if the purpose is to analyze the non-linear hydrological time series at a specific outlet that is naturally or artificially regulated, the LSTM model that remembers long-term information based on big data and reflects it on forecast after adjusting it is expected to be utilizable and applicable in various hydrological time series analysis.
주제어
#수문시계열 예측 딥러닝 지도학습 LSTM 모형
※ AI-Helper는 부적절한 답변을 할 수 있습니다.