Investigating the Ability of Artificial Neural Network (ANN) Models to Estimate Missing Rain-gauge Data

The correct forecasting of unrecorded data could be enormously helpful in designing water projects and preventing related damages. The conventional methods available for rainfall estimation usually take a long time to estimate the missing data, and their estimations may have many errors in the long-term simulations. In this study, the capabilities of different Artificial Neural Networks (ANNs) were analyzed in estimating missing data from the Ardabel plain rain gauge stations located in northwestern Iran. Accordingly, six different structures of ANNs were used, and their efficiencies in terms of the mean squared error, training, and validation determination coefficients to select better-estimated missing data were examined. The results revealed that the best model is composed of the feed-forward networks, trained by the Levenberg-Marquardt algorithm and considering only one hidden layer. For each of the stations with a complete data set, an ANN was trained. Data gaps from other stations were obtained by the proposed ANN models. Furthermore, an integrated ANN was developed to investigate the hidden spatial relationships among the rainfall data of the stations as well as temporal auto-correlations. The results indicated the superiority of the proposed integrated model. After the estimation of the rain data gaps, the K-means clustering method was also employed as a data pre-processing method to improve the accuracy of the estimation, and the method led to better results.