1. ABSTRACTS General Structure Background and Objective Downscaling CGCM climate change output scenario using the Artificial Neural Network model Kang Boosik 1 / Yang Jeong-Seok 2 1 Department of Civil & Environmental Engineering, Dankook University / 2 School of Civil & Environmental Engineering, Kookmin University Correspondence : Kang Boosik, Ph.D./Professor Department of Civil & Environmental Engineering, Dankook University #126 Jukjeon, Suji, Yongin, Korea 448-701 e-mail : bskang@dankook.ac.kr This study carried out the prediction of basin-wide climate change using GCM(Global Climate Model) climate change outlook scenario. To regionalize the original GCM scenario, the Artificial Neural Network (ANN) model was used. The 22 GCM output variables including precipitation flux, air pressure at sea level, near-surface daily-mean air temperature, surface upward latent heat flux etc, were used for potential predictor variables. The precipitation and temperature variables were used for predictands. The original GCM data is the CGCM3.1/T63 20C3M scenario (reference scenario) provided by CCCma (Canadian Centre for Climate Modeling and Analysis). The ANN learning process was performed from January 1997 to December 2000. The suggested ANN has a 3-layer perceptron (multi-layer perceptron; MLP) and back-propagation learning algorithm. The ANN predictors selected through the sensitivity analysis were utilized for final ANN model of Soyang and Chungju dam basin. Daily temperature and precipitation trend from 2001 to 2100 were suggested. The basin-wide prediction data of climate change scenario can be served as input data of long-term runoff model and give estimate of future available water resources. Keywords: Artificial Neural Network, Global Climate Model, multi-layer perceptron This study carried out the prediction of basin-wide climate change using GCM(Global Climate Model) climate change outlook scenario. To regionalize the original GCM scenario, the Artificial Neural Network (ANN) model was used. The 22 GCM output variables including precipitation flux, air pressure at sea level, near-surface daily-mean air temperature, surface upward latent heat flux etc, were used for potential predictor variables. The precipitation and temperature variables were used for predictands. The original GCM data is the CGCM3.1/T63 20C3M scenario (reference scenario) provided by CCCma (Canadian Centre for Climate Modeling and Analysis). The ANN learning process was performed from January 1997 to December 2000. The suggested ANN has a 3-layer perceptron (multi-layer perceptron; MLP) and back-propagation learning algorithm. The ANN predictors selected through the sensitivity analysis were utilized for final ANN model of Soyang and Chungju dam basin. Daily temperature and precipitation trend from 2001 to 2100 were suggested. The basin-wide prediction data of climate change scenario can be served as input data of long-term runoff model and give estimate of future available water resources. Keywords: Artificial Neural Network, Global Climate Model, multi-layer perceptron In order to reproduce regional climate information from the large-scale GCM outputs, the Model Output Statistics (MOS) post-processing based on multiple regression correlations between the predictand and available predictors is one of the useful methodologies. A number of other nonlinear techniques can be used to post-process NWP outputs, including generalized additive models (Vislocky and Fritsch, 1995), self-learning algorithms (e.g. Abdel-Aal and Elhadidy, 1995), and models based on artificial neural networks (ANN). ANN models have been proving to be useful in the field of hydrology (e.g. Silverman and Dracup, 2000). ANN models are capable of approximating extremely complex functions, while easier to apply than more traditional nonlinear statistical methods. This study carried out the prediction of basin-wide climate change using GCM(Global Climate Model) climate change outlook scenario. To regionalize the original GCM scenario, the Artificial Neural Network (ANN) model was used. One of the most popular network architectures is perhaps the multilayer perceptron, consisting of an input layer, one or more hidden layers, and an output layer. The input layer simply introduces the values of the input variables, while the hidden and output layer neurons are each connected to all of the units in the preceding layer. A diagram of a three-layer neural network is shown in Figure 1. The input layer consists of n I units, each of which receives one of the input variables. The so-called hidden layer is composed by n H units. Finally, the output layer consists of n K units, each of which computes a desired output (for the present study, only one output is needed). The neural networks developed for this study have a feed- forward structure: signals flow forward from input neuron through any hidden units, eventually reaching the output neurons. The mathematical equation of a three-layer NN can be written as: Study Area  Watershed : Soyang Dam  Location : Kangwon Province, Korea  # of sub-basin : 2 (#1011, #1012)  Area : 2,783.26 ㎢ (#1001 : 931.22 ㎢, #1002 : 1852.04 ㎢ )  # of Automatic Weather Station (AWS) : 8 stations  Data used for ANN training : daily temperature & precipitation for 1997 – 2000 (4 years)  Watershed : Soyang Dam  Location : Kangwon Province, Korea  # of sub-basin : 2 (#1011, #1012)  Area : 2,783.26 ㎢ (#1001 : 931.22 ㎢, #1002 : 1852.04 ㎢ )  # of Automatic Weather Station (AWS) : 8 stations  Data used for ANN training : daily temperature & precipitation for 1997 – 2000 (4 years) Schematic of a 3-layer neural network Conclusion Results Basin-wide daily temperature and precipitation projection from 2001 to 2100 were suggested for Soyang dam basin based on SRES B1 scenario downscaled from CCCma’s CGCM3.1/T63 output. When comparing 2081-2100 with 2001-2020, the winter and summer temperatures will be expected to increase 1.06 ℃ and 0.39 ℃, respectively, which shows warming will be more significant during winter season. T he winter and summer precipitation will be expected to increase 6.2mm and 21.2mm, respectively. To be more reliable results, the typhoon effects needs to be considered in the future. 33 rd IAHR Congress, Water Engineering for a Sustainable Environment, Vancouver, British Columbia, Canada, August 9-14, 2009 Where: ·x i (t) is the input to unit i of the input layer and z k (t) is the output obtained at unit k of the output layer ·i = 1, 2, …, n I where n I is the number of inputs ·h = 1, 2, …, n H where n H is the number of hidden units ·k = 1, 2, …, n K where n K is the number of outputs ·w h hi are the parameters, or weights, controlling the strength of the connection between the input unit i and the hidden unit h ·q i and q h are the thresholds · w o kh are the parameters controlling the strength of the connection between the hidden unit h and the output unit k ·f is the activation or transfer function: GCM scenario  Model : CGCM 3.1/T63 provided by CCCma  Grid Resolution : - Lateral : lon 2.81° × lat 2.81° (128 ×64 Gaussian grids) - Vertical : 63 layers  Available scenarios : SRES A1B, B1, A2  Reference scenario : CGCM3.1/T63 20C3M Scenario  Projection used in this study : CGCM3.1/T63 SRES B1 Scenario  Model output : - Monthly : 2001-2100 (2D, 3D) - Daily : 2001-2100 (2D) 2046-2065 (3D) 2081-2100 (3D)  Model : CGCM 3.1/T63 provided by CCCma  Grid Resolution : - Lateral : lon 2.81° × lat 2.81° (128 ×64 Gaussian grids) - Vertical : 63 layers  Available scenarios : SRES A1B, B1, A2  Reference scenario : CGCM3.1/T63 20C3M Scenario  Projection used in this study : CGCM3.1/T63 SRES B1 Scenario  Model output : - Monthly : 2001-2100 (2D, 3D) - Daily : 2001-2100 (2D) 2046-2065 (3D) 2081-2100 (3D) ANN Training with Reference Scenario 100-yr Projection under B1 Scenario 1011 sub-basin 1012 sub-basin Sub-basinMERRCORR 10110.520.65 10120.570.74 1011 sub-basin 1012 sub-basin DJF(Winter) JJA(Summer) Temperature Precipitation DJF(Winter) JJA(Summer) Sub-basinMERRCORR 1011-1.360.90 1012-0.780.88 Temperature Precipitation Abdel-Aal, R.E. and Elhadidy, M.A. (1995). Modeling and forecasting the maximum temperature using abductive machine learning. Weather and Forecasting 10, pp. 310-325 Silverman, D. and Dracup, J. (2000). Artificial Neural Networks and Long Range Precipitation Prediction in California, Journal of Applied Meteorology, 31:1, pp.57-66 Vislocky, R. L. and J. M. Fritsch (1995). Improved model output statistics forecasts through model consensus. Bull. Amer. Meteor. Soc., 76, pp.1157–1164. 10111012Mean DJFJJADJFJJADJFJJA 2001 – 20201.4619.710.3920.16 0.92519.9 2041 – 20601.8320.121.2520.58 1.5420.4 2081 - 21002.2120.101.7620.55 1.98520.3 10111012Mean DJFJJADJFJJADJFJJA 2001 – 20209.32276.0816.09234.75 12.71255.4 2041 – 206013.22293.0520.41251.12 16.82272.1 2081 - 210015.07295.9022.79257.29 18.93276.6 References TOPIC C: Water Engineering for the Protection and Enhancement of Natural Watershed and Aquifer Environments Technical Session 2, Track C-5: Climate Influences on Water Flow in Watersheds (August 10, Regency & Plaza Foyers, Balmoral)