A Time Series Reconstruction of Precipitation Records Using Artificial Neural Networks

Abstract

Feedforward multilayer neural networks have been used for the estimation of precipitation in selected rainfall collecting stations in Cyprus. Archived data collected for nine years and six control stations distributed around a target station have been used for training a suitable artificial neural network. Different neural network architectures and learning rates were tested, aiming at establishing a network that resulted in the best reconstruction of missing rainfall records. A multiple hidden layer architecture was chosen. This kind of architecture has been adopted for solving problems with similar requirements. The parameters used for the training of the network were collected at each control station. These are the Julian day, height, distance between target and control stations, and the precipitation. The correlation coefficient obtained for the training data set was 0.933. The verification of the network was done by using unknown data for the target station. This was done for a year, whose data were excluded from the training set. The correlation coefficient for the unknown case was 0.961. The prediction error was confined to less than 17.1mm of precipitation which is considered as acceptable.