HR: 1340h
AN: H13H-1679    [Abstracts]
TI: Construction of Evolutionary Artificial Neural Networks and its Application to Streamflow Forecasting
AU: Chang, F
EM: changfj@ntu.edu.tw
AF: National Taiwan University, No.1,Sec.4,Roosevelt Road, Taiwan(R.O.C.), Taipei, 10617,
AU: * Chen, Y
EM: wesleychen@wra.gov.tw
AF: National Taiwan University, No.1,Sec.4,Roosevelt Road, Taiwan(R.O.C.), Taipei, 10617,
AB: Selection of architecture for an artificial neural network (ANN) has significant influence on its successful application to the tasks to be performed. However, the conventional way for an ANN to process the optimization search is first to predefine fixed architecture and learning rule and then experience a series of trials and errors. This way usually leads the search to a local optimum and lack of efficiency and robustness. To improve the drawbacks of the conventional optimal process, this study introduces a novel algorithm, evolving artificial neural networks. With a hybrid encoding of network architecture, we apply genetic algorithm to optimize the encoded parameters of feedforward ANNsˇ¦ architecture. Then we optimize the connection weights of neurons by the scaled conjugate gradient algorithm. The terrific performance for forecasting of Mackey-Glass chaotic time series shows that the proposed algorithm concurrently possesses efficiency, effectiveness, and robustness. Besides, the application to the forecasting of 10-day reservoir inflows reveals again the algorithmˇ¦s excellent efficiency and robustness and its effectiveness which is superior to that of the AR(1) and ARMAX models. Keywords: Evolutionary artificial neural network (EANN), Genetic algorithm (GA), Hydrological systems, Forecasting, Reservoir inflow
DE: 1800 HYDROLOGY
DE: 1805 Computational hydrology
DE: 1846 Model calibration (3333)
DE: 1847 Modeling
DE: 1880 Water management (6334)
SC: Hydrology [H]
MN: 2007 Fall Meeting