Extreme Heat Identification with High Spatio-Temporal Resolution Using the Analog Ensemble Technique
Abstract
This work presents a downscaling algorithm based on Analog Ensemble (AnEn) method in order to identify Extreme Heat Events from datasets that vary in spatio-temporal resolution. The algorithm is tested using forecasts from Global Forecasts System (GFS) and observations from Weather Underground (WU) a for 2016-2017 over Manhattan, NY. GFS provides forecast every 3 hours at 28 km resolution whereas the WU data are interpolated every 15 minutes for 105 personal weather stations.
Problem arises because Rare Events (RE) are inherently difficult to capture using analog methods due that the analog bases future predictions on similarity of past occurrences. A Bias Correction (BC) technique is then used to compensate for the lack of past RE and reduces conditional negative bias introduced naturally by the AnEn. The difference between the current forecast and the average of the most similar past forecast is computed for a single variable. In the case of extreme events, this error will be large, as very few past forecasts will be similar to the current rare event. This difference is then added to each of the analog members, shifting the distribution by a fixed amount. Improved performances generated by BC are statistically validated using CRPS that shows lower values compared to non-bias corrected data. Given that AnEn predictions are generated independently at each location and Forecasts Lead Times (FLT), the members are not ordered causing a lack of consistency in space and time. A statistical reordering technique called Schaake Shuffle (SS) is applied to ensemble members to recover the spatio-temporal correlation. AnEn members for each day, location, and FLT are sorted according to the order of N observations, randomly selected from the training days, where N is equal to the number of AnEn members. Initial results show that AnEn method is able to generate accurate and well calibrated forecasts for both general and extreme events. This method can be further improved by running higher order regressions to capture non-linear relationship between current and past forecasts.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.A34C..06C
- Keywords:
-
- 3329 Mesoscale meteorology;
- ATMOSPHERIC PROCESSES;
- 3354 Precipitation;
- ATMOSPHERIC PROCESSES;
- 3355 Regional modeling;
- ATMOSPHERIC PROCESSES;
- 4313 Extreme events;
- NATURAL HAZARDS