Assessing the performance of solar radiation computing models and model selection procedures

This work has two main objectives. First, the redundancy of statistical indicators is analyzed. Sets of statistical indicators are prepared and their redundancy is analyzed. Selection procedures and model hierarchies are discussed. Statistical indicators based on errors have to be preferred instead of indicators based on relative errors. Minimal sub-sets of statistical indicators may be defined. Two sub-sets of indicators are recommended, i.e. (i) Mean Bias Error, Mean Absolute Error and the slope s of the best-fit line and (ii) Mean Bias Error, Root Mean Square Error and s. The t-statistics and Willmott's index of agreement may be added to these sets. Second, several procedures for models selection are analyzed. Different selection procedures and/or different input databases yield different hierarchies among models of comparable performance. The problem of the “best model” seems to have no solution. A reasonable approach is to classify models in “good” and other lower performance categories.