Gap Filling in Historical Data of Daily Precipitation in Rio Grande do Sul

Abstract

The gap filling in time series of precipitation is an important process for applications in hydrology, aiming the use of long series, avoiding that they are discarded. Thus, this study had as objective to gap filling in historical series of daily precipitation in Rio Grande do Sul, assisting in the use of these data in studies that require long term analysis. For that, we used historical series of 287 stations in the period between 1987 and 2016 and comparing the Multiple Linear Regression (MLR) and Artificial Neural Networks (ANN) methods, comparing to evaluating the filled values. An algorithm was developed to perform the following operations: i) identify the days with missing datas in each station; ii) identify the stations that can be used to fill each missing data; iii) identify all combinations of input to fill in failures in each station; iv) perform the adjustment/training of the MLR and ANN models; v) perform validation of the models based on the period without failure of each station. The main results indicate that the higher density of rainfall stations favors the process of filling of faults in historical series of precipitation, improving the quality of the filled series. Gap filling showed a higher coefficient of determination and lower mean absolute error using the MLR model in relation to RNA, possibly due to the strong linear correlation of the precipitation data of each site in relation to their neighborhood. The MLR model presented an average coefficient of determination (R²) of 0.697, while the RNA model obtained an average of 0.675. Considering the analysis by means of the mean absolute error (MAE), the mean values were 2.27 mm for MLR, while for RNA the error was 2.31 mm. It is concluded, considering the daily rainfall data set of RS, that there was a slight superiority of the RLM method in relation to RNA.

Keywords:
multiple regression; artificial neural networks; rainfall stations