Comparison of neural network methods for infilling missing daily weather records

Accurate estimate of missing daily precipitation data remains a difficult task particularly for large watersheds with coarse rain gauge network. Reliable and representative precipitation time series are essential for any rainfall–runoff model calibration as well as for setting-up any downscaling model for hydrologic impact study of climate change. This study investigates six different types of artificial neural networks namely the multilayer perceptron (MLP) network and its variations (the time-lagged feedforward network (TLFN)), the generalized radial basis function (RBF) network, the recurrent neural network (RNN) and its variations (the time delay recurrent neural network (TDRNN)), and the counterpropagation fuzzy-neural network (CFNN) along with different optimization methods for infilling missing daily total precipitation records and daily extreme temperature series. Daily precipitation and temperature records from 15 weather stations located within the Gatineau watershed in northeastern Canada, are used to evaluate the accuracy of the different models for infilling data gaps of daily precipitation and daily extreme temperatures. The experiment results suggest that the MLP, the TLFN and the CFNN can provide the most accurate estimates of the missing precipitation values. However, overall, the MLP appears the most effective at infilling missing daily precipitation values. Furthermore, the MLP also appears the most suitable for infilling missing daily maximum and minimum temperature values. The CFNN is similar to the MLP at infilling missing daily maximum temperature, however, it is less effective at estimating minimum temperature. The experiment results show that the dynamically driven networks (RNN and TDRNN) are the less suitable for infilling both the daily precipitation and the extreme temperature records, whereas the RBF appears fairly suitable only for estimating maximum and minimum temperature.