EarlyLEAD: A WRF ensemble demonstrating a data mining capability

Two key elements of the Linked Environments for Atmospheric Discovery (LEAD) project are dynamic adaptability and ensemble forecasting. In LEAD, dynamic adaptability is the notion of performing meteorological analysis and forecasting on demand in response to the weather. EarlyLEAD is a proof-of-concept effort within the LEAD project to autonomically identify significant mesoscale features and run the WRF model to follow the development of these features. The purpose of EarlyLEAD is to demonstrate and evaluate developing instantiations of some LEAD technologies, and to bring a subset of LEAD capabilities into computing environments that are not likely to have authorization on the TeraGrid, even when LEAD is fully functional. In addition, by adapting tools from different sources and combining them to create a new set of products, EarlyLEAD makes extensible LEAD technologies for faculty and students to use now.

EarlyLEAD begins with the most recent NAM or WRF forecast. The output from this forecast is subjected to the Phenomena Extraction Algorithm (PEA). PEA is a general purpose algorithm to detect and extract geophysical phenomena in science datasets, which does not depend on any specific domain heuristics or target data. PEA is a specialization of the Intelligent Data Thinning (IDT) algorithm developed at the University of Alabama at Huntsville (UAH) for other applications, which employs a tree based decomposition technique to recursively divide the data into multiple sections and calculates an objective information measure for each data partition. If the objective measure for a given partition is greater than a user specified threshold, the algorithm continues dividing this partitioned data further. If the objective measure is less than the user threshold, or when the data cannot be portioned further, then the algorithm terminates that recursive path. PEA and IDT are part of the Algorithm Development and Mining (ADaM) toolset developed at UAH as a component of the LEAD Service-oriented architecture.

User-defined forecast quantities, such as the u and v components of wind velocity or omega (vertical velocity) and precipitation fields, are autonomically selected from the initial model output and subjected to the PEA. The algorithm identifies and outputs the center latitude and longitude of regions of interest (ROI), and prioritizes according to their significance. The PEA can be adjusted by the user to extract the ROI according to geographical location, pre-selected meteorological phenomena, and other user preferences. Once the locations of ROIs have been prioritized, the center latitude and longitude of the ROI with the highest significance is used as the basis for establishing a high resolution WRF model domain over that region. The WRF is initialized with the model (NAM or WRF) output, then run to produce a short-term forecast. The output from the WRF forecast is again subjected to the PEA to identify where to move the WRF domain. This sequential process of output-to-PEA-to-WRF-to-output steers the WRF domain to follow the interesting weather identified through the PEA. The PEA is being tested at Millersville and Howard Universities, and at Unidata Program Center (UPC). The three WRF products are then sent to the UPC LEAD test bed system where they are made available on the OPeNDAP server and cataloged using THREDDS. The UPC Integrated Data Viewer (IDV) is then used to compare and merge the ensemble.