An Exploration of the Analog Ensemble Search Space Extension and Spatio-Temporal Reconstruction.

The Analog Ensemble (AnEn) technique enables probabilistic predictions using fewer computational resources than traditional ensemble prediction methods. The technique performs well for wind and solar energy prediction, air quality forecasting, 2-m temperature, 10-m wind speed, and select downscaling efforts. The original AnEn Independent Search (IS) technique presented by Delle Monache et al. (2013) is implemented at a single point in space over a n-point time window and uses a historical repository of corresponding deterministic predictions and observations. Historical repositories of corresponding deterministic predictions and observations can be temporally limited or may not be available for experimentation. Therefore, the first component of this research determines a means of artificially expanding the historical search repository in order to foster an improved AnEn prediction. The technique developed is called the AnEn Search Space Extension (SSE). The AnEn SSE technique utilizes spatial information, which is available in abundance, to fill in for a lack of temporal information. For this investigation, deterministic Numerical Weather Prediction (NWP) data comes from the European Center for Medium-Range Weather Forecasts (ECMWF) high-resolution model which has a horizontal resolution of 0.25°, forecast lead times at 3-hour intervals. ECMWF operational analysis fields are used as the observational dataset and are available at 6-hour intervals. Data spans a three-year time period from January 2013 to December 2015. Results show that the SSE generally outperforms the AnEn IS and ECMWF prediction when short search spaces are used as the search space for a similar future season.

Both the AnEn IS and AnEn SSE enable predictions at specific points in space and provide improved probabilistic prediction using fewer computational resources. It is advantageous to have a prediction of a gridded probabilistic meteorological fields. However, when the AnEn IS and AnEn SSE predictions are generated using a grid-based implementation each result in spatio-temporal inconsistencies. Sperati et al. (2017) present a post-processing technique called the Schaake Shuffle (SS) in order to improve spatio-temporal consistency. Results from Sperati et al. (2017) showed the AnEn IS + SS provided an improved prediction over the original ECMWF prediction. However, the SS has limitations that require further investigation. This second component explores three topics: 1) Seeks to understand discrepancies between the SS-Clark versus the SS-Sperati implementations; 2) assesses the performance of the AnEn SSE using the different aforementioned implementations; 3) develops a deeper understanding of the continuity in space and time as reconstructed by the SS. Findings from this work have implications in fields where a probabilistic prediction is needed but historical datasets may be limited.