Comparison of Artificial Neural Network and Regression Models for Filling Temporal Gaps of Meteorological Variables Time Series

Continuous meteorological variable time series are highly demanded for various climate related studies. Five statistical models were tested for application of temporal gaps filling in time series of surface air pressure, air temperature, relative air humidity, incoming solar radiation, net radiation, and soil temperature. A bilayer artificial neural network, linear regression, linear regression with interactions, and the Gaussian process regression models with exponential and rational quadratic kernel were used to fill the gaps. Models were driven by continuous time series of meteorological variables from the ECMWF (European Centre for Medium-range Weather Forecasts) ERA5-Land reanalysis. Raw ECMWF ERA5-Land reanalysis data are not applicable for characterization of specific local weather conditions. The linear correlation coefficients (CC) between ERA5-Land data and in situ observations vary from 0.61 (for wind direction) to 0.99 (for atmospheric pressure). The mean difference is high and estimated at 3.2 °C for air temperature and 3.5 hPa for atmospheric pressure. The normalized root-mean-square error (NRMSE) is 5–13%, except for wind direction (NRMSE = 49%). The linear bias correction of ERA5-Land data improves matching between the local and reanalysis data for all meteorological variables. The Gaussian process regression model with an exponential kernel based or bilayered artificial neural network trained on ERA5-Land data significantly shifts raw ERA5-Land data toward the observed values. The NRMSE values reduce to 2–11% for all variables, except wind direction (NRMSE = 22%). CC for the model is above 0.87, except for wind characteristics. The suggested model calibrated against in situ observations can be applied for gap-filling of time series of meteorological variables.