Deep learning (DL) models have become immensely popular in recent years, with many models creating accurate and high-skill predictions for a wide range of atmospheric phenomena. Using DL models for predicting convection and associated hazards has experienced some of the most substantial gains in skill. The National Severe Storms Laboratory (NSSL) has created the experimental Warn-On-Forecast System (WoFS) to increase warning lead times through probabilistic short-term forecasts of individual thunderstorms. Currently, the WoFS has a shortcoming of missing storms due largely to poorly initialized environments. To help mitigate this issue, we developed a U-Net deep learning model to predict locations of thunderstorms trained on WoFS model data consisting of environmental data, such as CAPE and CIN, and intra-storm variables, such as WoFS ensemble average, mean, and max composite reflectivity and updraft and downdraft velocities. To address the issue of poorly initialized environments and lagging data assimilation, the model also has access to Multi-Radar/Multi-Sensor System (MRMS) data valid at the WoFS initialization is also used as an input. To evaluate the skill of the DL-based guidance, different baseline methods were tested to ensure a substantial performance increase. Comparing the performances of the WoFS baseline and DL model on an independent testing dataset, we were able to increase the maximum critical success index from 0.17 to 0.27, along with increasing the reliability and discrimination of the predictions. Using MRMS composite reflectivity proved to be vital for the DL model's performance when predicting values >= 40 dBZ. Through this work, we demonstrate DL models are an effective and efficient solution to improving the skill of the WoFS forecast of convection with a 30-minute lead time.