P.Keith Kelly, Haonan Chen, and Mohit Kumar
Accurate nowcasting (short-term prediction, 0–6 h lead time) of high-impact weather such as extreme convective storms plays a critical role in natural disaster monitoring and mitigation. A number of nowcasting approaches have been developed in the past two decades, including optical flow and the tracking radar echoes by correlation (TREC) system, which are still in use in operational environments. These techniques are based on radar echo map extrapolation that determines the velocity and direction of precipitation systems using historical and current radar observations. However, the skill of these traditional extrapolation methods decreases rapidly within the first hour. Recent research has demonstrated the appealing performance of artificial intelligence (AI) and deep learning for radar-based nowcasting of high-impact weather. But none of the previous studies has applied deep learning-based nowcasting for phased array radar observations, which can provide more insights into the precipitation systems through enhanced spatiotemporal resolution compared to conventional weather radars. This study presents a self-attention-based recurrent neural network (SaGRU) to enhance radar-based nowcasting using phased array radar observations. The SaGRU model is trained and tested with multiple precipitation events in Colorado. In addition, this research quantifies the impact of resolution and the sequence length of input radar observations on the nowcasting performance, which further highlights that the high-resolution phased array radar system can enhance precipitation observations and prediction.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner