Application of ANN to the prediction of missing daily precipitation records, and comparison against linear methodologies
Centro de Cálculo - Facultad de Ingeniería
Universidad de la República,
Depending upon the user, weather records can be used as they are, or they need to be imputated prior its use. Despite the fact that general methods for meteorological variables exist, they are difficult to apply for daily rain. A specially difficult feature is that the overwhelming majority of the records (>80%) are of zero rain, leading to a very non-gaussian distribution. Other characteristic is the low autocorrelation of the time series.
The test region was the Santa Lucia river catchment area of 13000 km2, at 35°S near the Atlantic; its yearly accumulated precipitation values are around 1000 mm, without a clear dry or wet season. The selected subset has 20 years long and 10 stations; 30% of the events show missing values.
A Monte Carlo simulation was designed, randomly choosing both date and station for the missing values and afterwards different imputation procedures were successively applied. Some statistics which characterize the distribution of the absolute error, namely its expected value, variance and 75, 85 and 95 percentile have been derived in order to compare the results.
Both traditional linear meteorological interpolation procedures as well as a suite of Backpropagation Artificial Neural Networks(ANN) has been compared. The present results are not very good, and show that is possible to imputate with a mean error of 2 mm/day and a RMS of 7 mm/day using both linear and nonlinear procedures, while ANN seems to be more robust against outliers.