Improving WoFS-PHI Watch to Warning Severe Weather Guidance through new Predictor and Target Datasets

Abstract

Warn-on-Forecast System-Probabilistic Hazard Information (WoFS-PHI) is a real-time machine learning algorithm that predicts individual severe weather hazards (hail, wind, and tornadoes) for up to 4 hours of lead time. The original version of WoFS-PHI used predictors from WoFS and ProbSevere Version 2 (PS2) to predict local storm reports (LSRs). While previous work found WoFS-PHI to be more skillful than machine learning baseline forecasts from WoFS or PS2 individually, an open question is whether using different predictor and/or target datasets would improve the skill of WoFS-PHI. In this thesis, multiple experiments are conducted to examine the (combined and individual) influence of: ProbSevere Version 3 (PS3) and Tornado Probability Algorithm (TORP) predictors, 1-h vs. 2-h time windows, and hazard-specific warning machine learning targets. New WoFS-PHI models were trained with varying sets of predictors and target datasets. Models were verified using Brier Skill Score, performance diagrams, and attributes diagrams. Predictor importance was analyzed through relative tree interpreter and accumulated local effect curves. Results show that including PS3 and TORP in the predictors modestly improves forecast skill for all three hazards, but especially for tornadoes. Predictor importance results showed that PS3 predictors were slightly more important to WoFS-PHI than PS2, and PS3 were less important for 2- than 1-hour forecasts regardless of lead time, but WoFS predictors were more important for 2- than 1-hour forecasts. TORP had relatively minimal impacts compared to WoFS and ProbSevere due to a lack of TORP objects in the training dataset. However, when TORP objects were present, they were quite influential to forecasts of all hazards. Results showed that BSSs for WoFS-PHI models trained on warnings and LSRs were much higher than BSSs for WoFS-PHI models trained on LSRs only. This suggests that WoFS-PHI is more skillful at predicting warnings relative to warning climatology than predicting LSRs relative to LSR climatology. In other words, warnings-and-LSRs can be a more predictable target for WoFS-PHI than LSRs alone. While additionally training on warnings offers both advantages and disadvantages, overall findings here suggest that warnings should at least be considered to supplement LSRs as a target for severe weather forecasting.