Zhe Xu, Aleksei Rutkovskii, Brenda Philips, and Michael Zink
Paros Center of Atmospheric Research
University of Massachusetts
Amherst, MA 01003
High-resolution Numerical Weather Prediction (NWP) models have significantly improved forecast accuracy, which is vital across sectors such as agriculture, urban planning, and disaster management. Traditional NWP models face several challenges in short-term precipitation forecasting. These include high computational demands, long model initialization times, and substantial financial long-term costs due to their reliance on extensive computational resources and supercomputing infrastructure (Ben Bouallègue et al. 2024; Kaufmann & Schubiger 2003; Sun et al. 2014). Additionally, these models may exhibit reduced accuracy over complex terrains and are constrained by limited grid resolutions, which can result in missing small-scale or localized weather phenomena (Kaufmann & Schubiger 2003; Sun et al. 2014).
In contrast, machine learning (ML) models offer several benefits that address these limitations. By learning patterns from historical data, ML models can generate forecasts quickly, reducing the need for extensive computational resources and lowering financial costs (Aksoy & Dahamsheh 2009; Ben Bouallègue et al. 2024). Among these, generative machine learning models represent a more capable and expressive field within ML, offering several advantages over traditional models. Generative models, when implemented using architectures like conditional generative adversarial networks (cGANs) and diffusion models, are adept at capturing the complex, nonlinear dynamics of weather systems.
State-of-the-art cGAN models, such as DGMR and NowcastNet, provide single prediction outcomes as well as express predictive uncertainties, crucial for understanding and preparing for extreme weather events (Ravuri et al. 2021; Zhang et al. 2023). Adversarial training and the inclusion of latent variables help these models balance between generated and real data, improving accuracy in spatial structures and temporal continuity, and enabling the generation of diverse prediction results (Arjovsky et al. 2017; Majumdar et al. 2021; Sun et al. 2014). These features highlight the advantages of generative machine learning models in capturing complex meteorological processes and providing rich information for meteorological decision-making.
In our study, we evaluate the performance and applicability of GANs for short-term rainfall rate prediction using X-band radar data from the CASA-DFW radar network in the Dallas-Fort Worth (DFW) region (Philips et al. 2017). The CASA DFW Network is a network of 4- 7 X-band radars in North Texas that have been generating high spatiotemporal resolution single radar moments and multi-radar products since 2016.
We focus on comparing several generative machine learning models, including DGMR and NowcastNet, to verify the claims made in their original publications. Specifically, DGMR claims to outperform PySTEPS, a widely-used inference-based solution for precipitation nowcasting, by delivering more accurate and realistic spatiotemporal predictions (Pulkkinen et al. 2019). NowcastNet further claims to surpass both DGMR and PySTEPS, offering improved nowcasts and maintaining higher accuracy, especially for extreme-precipitation events. Through this comparative analysis, we will present the results of our work regarding the models' capabilities, propose and validate improvements based on the findings, and enhance the accuracy and applicability of precipitation forecasting models, thereby advancing meteorological prediction technologies.
References
Aksoy, H., and A. Dahamsheh, 2009: Artificial neural network models for forecasting monthly precipitation in Jordan. Stochastic Environmental Research and Risk Assessment, 23 (7), 917–931, https://doi.org/10.1002/met.127.
Arjovsky, M., and Coauthors, 2017: Wasserstein generative adversarial networks. Proceedings of the 34th International Conference on Machine Learning, Vol. 70, 214–223, https://doi.org/10.48550/arXiv.1701.07875.
Ben Bouallègue, Z., and Coauthors, 2024: The Rise of Data-Driven Weather Forecasting: A First Statistical Assessment of Machine Learning–Based Weather Forecasts in an Operational-Like Context. Bull. Amer. Meteor. Soc., 105, E864–E883, https://doi.org/10.1175/BAMS-D-23-0162.1.
Kaufmann, P., and F. Schubiger, 2003: Precipitation forecasting by a mesoscale numerical weather prediction (NWP) model: Eight years of experience. Hydrology and Earth System Sciences, 7 (5), 812–832, https://doi.org/10.5194/hess-7-812-2003.
Majumdar, S. J., and Coauthors, 2021: Multiscale Forecasting of High-Impact Weather: Current Status and Future Challenges. Bull. Amer. Meteor. Soc., 102, E635–E659, https://doi.org/10.1175/BAMS-D-20-0111.1.
Philips, B., and Coauthors, 2017: Tracking tornados down streets: Using CASA radars in real-time severe weather warning operations in North Central Texas. IEEE International Geoscience and Remote Sensing Symposium, 2017, 5973–5976, https://doi.org/10.1109/IGARSS.2017.8128370.
Pulkkinen, S., and Coauthors, 2019: Pysteps: An open-source Python library for probabilistic precipitation nowcasting (v1.0). Geoscientific Model Development, 12 (10), 4185–4219, https://doi.org/10.5194/gmd-12-4185-2019.
Ravuri, S., and Coauthors, 2021: Skilful precipitation nowcasting using deep generative models of radar. Nature, 597 (7878), 672–677, https://doi.org/10.1038/s41586-021-03854-z.
Sun, J., and Coauthors, 2014: Use of NWP for Nowcasting Convective Precipitation: Recent Progress and Challenges. Bull. Amer. Meteor. Soc., 95, 409–426, https://doi.org/10.1175/BAMS-D-11-00263.1.
Zhang, Y., and Coauthors, 2023: Skilful nowcasting of extreme precipitation with NowcastNet. Nature, 619 (7970), 526–532, https://doi.org/10.1038/s41586-023-06184-4.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner