Using analog ensembles with alternative metrics for hindcasting with multistations

This study concerns making weather predictions for a location where no data is available, using meteorological datasets from nearby stations. The hindcast with multiple stations is performed with different variants of the Analog Ensemble (AnEn) method. In addition to the traditional Monache metric used to identify analogs in datasets from one or two stations, several new metrics are explored, namely cosine similarity, normalization, and k-means clustering. These were analyzed and benchmarked to find the ones that bring improvements. The best results were obtained with the k-means metric, yielding between 3% and 30% of lower quadratic error when compared against the Monache metric. Also, by making the predictors to include two stations, the performance of the hindcast improved, decreasing the error up to 16%, depending on the correlation between the predictor stations.