92nd American Meteorological Society Annual Meeting (January 22-26, 2012)

Tuesday, 24 January 2012: 12:00 AM
An Artificial Neural Network to Predict Thunderstorms in South Texas
Room 242 (New Orleans Convention Center )
Waylon G. Collins, NOAA/NWS, Corpus Christi, TX; and P. Tissot

A supervised, feedforward, multi-layer perceptron, Artificial Neural Network (ANN) model was developed to forecast thunderstorm occurrence over portions of a South Texas domain 3 to 12 hours in advance with temporal and spatial accuracies of 4-hours and 20-km, respectively. We refer to this model as the Thunderstorm Artificial Neural Network (TANN). The TANN framework involves the use of predictors that account for both the meso-γ scale (12-km grid spacing) environment conducive to convective development and sub-grid scale (4-km grid spacing) processes that contribute to convective initiation (CI). The environmental conditions conducive to CI were assessed based on selected output variables/parameters from a combination of the 12-km Eta and the 12-km Weather Research and Forecasting Non-hydrostatic Mesoscale Model (WRF-NMM) Numerical Weather Prediction (NWP) models. With regard to the smaller scale processes, we incorporate sub-grid scale soil moisture and soil moisture heterogeneity. Studies reveal that, in the absence of synoptic scale forcing, the exact location of CI is dependent on small scale regions of enhanced soil/ambient moisture and/or the convergent part of the surface wind pattern produced by soil moisture heterogeneity. The soil moisture proxy used in this study was output from the Antecedent Precipitation Index (API) model applied to the 4-km MPE (Multisensor Precipitation Estimator) output from the National Weather Service (NWS) West Gulf River Forecast Center (WGRFC). The data set for this study covered the period from April 2004 to December 2010. The TANN represents the post-processing of the foregoing environmental and sub-grid scale output. In particular, the TANN was trained and validated (using various architectures including 1-hidden layer, the logsig and purelin transfer functions, and both the levenberg-Marquardt and Bayesian regularization training algorithms from the MATLAB Neural Network Toolbox, and with various fractions of the data set divided between training, validation and testing), and then tested to assess its performance (ability to generalize.) Performance metrics included the Relative Operating Characteristic (ROC) curve, and scalar parameters such as the Heidke Skill Score (HSS), and the Odds Ratio Skill Score (Yule's Q). We assessed TANN model performance relative to observations (cloud-to-ground lightning from the National Lightning Data Network) and relative to the output from forecasters at the NWS Weather Forecast Office in Corpus Christi, Texas (WFO CRP). Preliminary results indicate that the TANN possess skill. The TANN represents an alternative to the paradigm of using of high resolution (≤4-km) NWP output to forecast thunderstorm occurrence.

Supplementary URL: