역학적으로 규모축소된 남한의 겨울철 기온에 대한 인공신경망과 다중선형회귀모형을 이용한 보정 비교 연구
역학적으로 규모축소된 남한의 겨울철 기온에 대한 인공신경망과 다중선형회귀모형을 이용한 보정 비교 연구
复制标题
我想知道我想说什么 我想知道我想知道什么
DOI：
--
复制
发表时间：
2005
影响因子：
2.3
通讯作者：
차유미
中科院分区：
文献类型：
--
作者：
안중배;차유미
In this study, the systematic and random errors generated by dynamic downscaling have been corrected using Artificial Neural Networks (ANN) and Multiple Linear Regression (MLR) and compared for the reproduction of the realistic regional climate over South Korea for the period of 1990-2001. That is, the 10-daily January surface air temperature over South Korea produced by the regional model, MM5, has been corrected by MLR based on linear assumption between predictors and observation and by ANN which considers nonlinear relationship between the two and then the two corrected results have been compared for the examination of the improvement of model result. The correlation analysis with observation shows that the ANN-corrected results have a better correlation than those of MLR, although the results of the two methods show the 99% confidence level in the training process. On the contrary, the uncorrected results do not show any significant correlation. In the cross-validation process, as same as training, the ANN-corrected results have a better correlation than those of MLR. The confidence level of the cross-validated results using MLR is lower than those of trained results. Root Mean Square Error(RMSE) distribution also shows that the ANN-corrected results have a smaller RMSE than those of MLR in two processes so that ANN-corrected results are closer to an observation. Both the corrected results using two methods reproduce more realistic time series than uncorrected results in two processes. Particularly, ANN is superior to predicting the extreme values such as maximum and minimum temperature. From the scatter plot analysis, the ANN-corrected results have the much higher linear relationship with the observation than those of MLR. Through the analyses above, it is concluded that the systematic and random errors in the dynamically downscaled regional climate model can be improved using correction methods, and that the ANN-corrected results show better correction than those of MLR due to nonlinearity of the model characteristics.
