J3.4 Use of an Artificial Neural Network to Forecast Thunderstorm Location: Performance Enhancement Attempts

A feed-forward, supervised, multi-layer perceptron Artificial Neural Network (ANN) was developed to test the following hypothesis: An ANN can be developed to successfully forecast thunderstorm activity up to 24 hours in advance, and with a spatial accuracy of 20-km, wherein ANN inputs include selected output from (1) deterministic mesoscale Numerical Weather Prediction (NWP) models, and from (2) selected sub-grid scale data that contributes to convective initiation, or CI (Collins and Tissot, 2007, hereafter CT07.) We are not aware of any other project involving the use of both NWP output and sub-grid scale data, as inputs into an ANN, with the desire to forecast thunderstorm activity with an accuracy of 20-km. The underlying logic of this novel hypothesis is that the NWP model output provides a forecast of whether the larger mesoscale environment is conducive to CI while the sub-grid scale data determines the extent to which convection could be triggered at a particular location. The ANN serves as a means to map the highly non-linear relationship between the foregoing inputs and thunderstorm occurrence; an ANN model to forecast thunderstorms would result. This represents a paradigm shift away from the sole use of high-resolution (horizontal grid spacing ≤ 4-km) NWP models to forecast thunderstorms, which, as suggested by Elmore et. al. (2002), may not be a reliable strategy. Results from CT07 were mixed: The model’s ability to forecast thunderstorm activity was encouraging, yet the number of false alarms was high. It was surmised that false alarms can be reduced by incorporating more relevant sub-grid scale data, and increasing the number of relevant NWP parameters that contribute to CI. This study represents such an attempt to improve the ANN’s performance. In Section 2, we discuss ANNs in a general sense. In Section 3, we describe the framework used to develop this ANN model. Section 4 contains a detailed description of the specific data inputs. In particular, we discuss how each parameter is related to thunderstorm development, and the specific data processing methods. In the final sections, the results (Section 5) and discussion and conclusions (Section 6) are presented. Portions of sections 2 and 4 contain information incorporated from our earlier study (CT07.)