A comparative study of imputation methods for improving statistical post-processing of weather forecasts

Accurate weather forecasting is vital for societal decision-making in sectors such as renewable energy, agriculture, and disaster management. Statistical post-processing techniques play a critical role in calibrating forecasts and addressing issues of model bias and ensemble dispersion. However, many post-processing methods rely on complete and high-quality datasets, and the presence of missing data can significantly undermine their effectiveness. This study presents a comparative analysis of imputation methods aimed at bridging data gaps to enhance the performance of statistical post-processing techniques.

The evaluation process focuses on a selection of widely used imputation approaches, including ensemble member mean substitution, persistence, Fourier fit, and Neural Networks. These methods are assessed using the forecasts and observations from the EUPPBench dataset by introducing randomly selected missing data, focusing on metrics such as imputation accuracy and their impact on post-processing performance. To quantify the benefit of missing data imputation the study compares different post-processing techniques, ranging from the simpler EMOS to the more advanced Neural Networks, where the latter is known to be more affected by incomplete data.