Imputation of missing values in spatio‐temporal solar radiation data

Solar radiation was measured almost every 10 minutes for 2 years at ten sites in the Pentland Hills to the south of Edinburgh. A few values were missing because of instrument failure or because the data logger was occupied in data transfer. Two ad hoc methods for replacing the missing values are compared by cross‐validation. It is found to be better to interpolate sequences of up to four missing values, and the first and last missing values in longer sequences, using data from the same site, but otherwise it is better to average synchronous data from the other nine sites. A spatio‐temporal model of the second‐order moments of the data is identified and used to specify optimal linear predictors based on both within‐ and between‐site data.