EMOS is a popular machine learning model used in meteorology for post-processing ensemble weather forecasts. In our work we use air temperature forecasts of an 11-member ensemble to train the EMOS model and investigate how variations in training data influence its performance. We manipulate training data by introducing sliding and fixed window approaches and by variating their length. We also use regularization and add climatological data. We evaluate the model performance using CRPS, RMSE and MAE metrics and conclude that the post-processing with the best EMOS model configuration is on average 23,97% better than the performance of the raw ensemble.
|
