신경망모형들의 학습알고리즘을 이용한 홍수량 예측 비교 (Comparison of Runoff Prediction Using Neural Network Training Algorithm)

서지정보

· 발행기관 : 한국수처리학회
· 수록지 정보 : 한국수처리학회지 / 18권 / 4호 / 79 ~ 89페이지
· 저자명 : 차호섭, 고근윤, 최윤영

초록

The purpose of this study is to improve our flood forecasting capability using a multiple neural network Algorithm. The Artificial Neural Network (ANN), Generalized Regression Neural Network (GRNN), and the Neuro-Fuzzy Model (NFM) theory were used during this study. The use of Geographic Information Systems (GIS) enabled us to analyze spatial modification for the study areas. These models are capable of the mathematical interpretation of the human thought process and are useful for increasing the level of output of the models and reducing the uncertainty of model parameters that occurs from the modeling process for the forecast of the river runoff. The algorithm utilizes a multi-layer neural network to develop a river runoff forecasting model. The number of hidden layer nodes, considering the number of input layer nodes, involve 11 models by the prior forecasting hour, that is one, two and four hours ahead (input layer node standard). The forecasting result was simultaneously checked utilizing both the diagram method and the statistical method. Four actual flood events that occurred between 1997 and 2003 were selected for each proposed model for learning for neural network algorithm. Among all developed models, three ANNs, four GRNNs, and three NFMs, one of NFM models, NFM-4-4 model, showed the best performance forecasting flood runoff in one, two, and four hour increments in June 2002 flood events. The comparison between the ANN and the NFM showed that the NFM provided a more accurate peak rate for all runoff forecasts lasting up to four hours.

영어초록

The purpose of this study is to improve our flood forecasting capability using a multiple neural network Algorithm. The Artificial Neural Network (ANN), Generalized Regression Neural Network (GRNN), and the Neuro-Fuzzy Model (NFM) theory were used during this study. The use of Geographic Information Systems (GIS) enabled us to analyze spatial modification for the study areas. These models are capable of the mathematical interpretation of the human thought process and are useful for increasing the level of output of the models and reducing the uncertainty of model parameters that occurs from the modeling process for the forecast of the river runoff. The algorithm utilizes a multi-layer neural network to develop a river runoff forecasting model. The number of hidden layer nodes, considering the number of input layer nodes, involve 11 models by the prior forecasting hour, that is one, two and four hours ahead (input layer node standard). The forecasting result was simultaneously checked utilizing both the diagram method and the statistical method. Four actual flood events that occurred between 1997 and 2003 were selected for each proposed model for learning for neural network algorithm. Among all developed models, three ANNs, four GRNNs, and three NFMs, one of NFM models, NFM-4-4 model, showed the best performance forecasting flood runoff in one, two, and four hour increments in June 2002 flood events. The comparison between the ANN and the NFM showed that the NFM provided a more accurate peak rate for all runoff forecasts lasting up to four hours.

참고자료

· 없음