Application of Adaptive Variable Structure of ANN to Distributed Rainfall Interpolation

Abstract

Rainfall interpolation is a hot issue in the study of distributed hydrological model due to its complexity. For the rainfall interpolation methodology, Artificial Neural Network (ANN) is more excellent in both precision and efficiency than those traditional methods. Furthermore, with the purpose of reducing uncertainty of hidden layers in ANN, this paper constructs a Back Propagation Artificial Neural Network (BPANN) model based on adaptive variable number of hidden layer’s nodes to estimate the rainfall in Hubei province. Result proves that the method of BPANN has better performance than conventional interpolation method such as Inverse Distance Weight Method (IDWM), the Mean Relative Error (MRE) of BPANN is 20.98%, whereas the MRE of IDWM is 37.57%. The result also shows that it is optimum for river basin of Hubei province when structure of BPANN model is 3-12-1, and the MRE is 19.57%.