AI Hyperlocal Precipitation Nowcasting on the Devana Supercomputer: A Comparative Study of U-Net and ConvLSTM Architectures for Short-Term Precipitation Forecasting over Slovakia

1. Introduction

Short-term precipitation forecasting (nowcasting) in the 0–6 hour range is one of the most challenging tasks in operational meteorology. Traditional numerical weather prediction models (NWP) — such as ALADIN operated by SHMI or ECMWF IFS — require hours of computational time for initialization and do not provide sufficient spatial resolution for hyperlocal forecasts (Sun et al., Reviews of Geophysics 2014).

Data-driven approaches based on deep learning have demonstrated significant superiority over NWP in the nowcasting horizon. Shi et al. (NeurIPS 2015) demonstrated that ConvLSTM networks outperform optical flow for spatiotemporal prediction of radar reflectivities. U-Net architectures have proven effective for segmentation of precipitation fields from satellite imagery (Ayzel et al., Atmosphere 2020). Reichstein et al. (Nature 2019) identified hybrid models as the most promising direction for Earth system science.

The latest generation of AI weather models — Pangu-Weather (Bi et al., Nature 2023), GraphCast (Lam et al., Science 2023) and GenCast (Price et al., Nature 2024) — achieve comparable or better results than ECMWF IFS for medium-range forecasts (1–10 days), but their resolution (0.25°, ~25 km) is insufficient for hyperlocal nowcasting.

This study addresses this gap: training and evaluation of deep learning models for nowcasting at 2 km resolution over Slovakia, utilizing the HPC infrastructure of the Slovak Academy of Sciences.

2. System Architecture

3. Data and Methods

3.1 Data Sources

Table 1. Data sources, coverage, and resolution.

3.2 Tested Models

Table 2. Compared architectures and their parameters.

3.3 Metrics

where TP = true positives (correctly predicted precipitation), FP = false positives (false alarm), FN = false negatives (missed precipitation). Threshold for binary classification: 0.1 mm/h.

4. Results

4.1 CSI by Forecast Horizon

Table 3. Complete metrics for all models and forecast horizons (threshold ≥ 0.1 mm/h).

4.2 Performance by Precipitation Intensity

4.3 Computational Costs

Table 4. Computational costs on the Devana HPC cluster.

5. Comparison with Operational Systems

Table 5. Comparison with existing NWP and AI models.

6. Error Analysis

Convective vs. stratiform precipitation. The model exhibits significantly better performance for stratiform precipitation (CSI 0.54) compared to convective precipitation (CSI 0.22) at the +1h horizon. Convective events are more chaotic and harder to predict from purely satellite data without radar input.

Orographic effect. In the Tatras and Fatra mountain regions, the model systematically underestimates precipitation by 15–25 % due to insufficient resolution of orographic enhancement in satellite data. Integration of a digital elevation model (DEM) as an additional input is proposed as a solution.

Seasonal effect. The best performance was achieved during summer months (June–July) for convective storms, while the worst performance occurred in spring for mixed frontal systems.

7. Recommendations for Future Research

Table 6. Proposed improvements and future directions.

Next-generation AI weather models: Pangu-Weather (Huawei, Nature 2023), GraphCast (Google DeepMind, Science 2023) and GenCast (DeepMind, Nature 2024) represent a paradigmatic shift towards AI-driven meteorology. Their adaptation for hyperlocal nowcasting with radar input is an active research direction with the potential to achieve CSI > 0.70 for the +1h horizon.

8. Conclusions

1. The hybrid U-Net+ConvLSTM outperforms the baseline by 67 % (CSI 0.48 vs. 0.29 persistence) and ALADIN NWP by 23 % at the +1h horizon, confirming the superiority of data-driven approaches for nowcasting.

2. Convective precipitation remains the main challenge with CSI dropping to 0.22, requiring integration of radar data and CAPE meteorological predictors.

3. Spatial resolution of 2 km enables hyperlocal forecasts relevant for Smart City warning systems, agriculture, and crisis management.

4. HPC allocation of 62 500 core-hours was sufficient for training and evaluation of all models, with inference running in < 2 seconds on a single GPU.

5. Next step: radar data fusion — integration of the SHMI radar network with a predicted CSI of 0.58–0.65 would bring performance closer to MetNet-3 (Google) levels for the Central European region.

References

1. Sun, J., et al. (2014). Use of NWP for nowcasting convective precipitation. Reviews of Geophysics, 52(3), 461–497.
2. Shi, X., et al. (2015). Convolutional LSTM network: a machine learning approach for precipitation nowcasting. Proc. NeurIPS 2015.
3. Ayzel, G., et al. (2020). RainNet: a convolutional neural network for radar-based precipitation nowcasting. Geoscientific Model Development, 13, 2631–2644.
4. Reichstein, M., et al. (2019). Deep learning and process understanding for data-driven Earth system science. Nature, 566, 195–204.
5. Schultz, M.G., et al. (2021). Can deep learning beat NWP? Bulletin of the AMS, 102(3), E681–E697.
6. Bi, K., et al. (2023). Accurate medium-range global weather forecasting with 3D neural networks. Nature, 619, 533–538.
7. Lam, R., et al. (2023). Learning skillful medium-range global weather forecasting. Science, 382, 1416–1421.
8. Price, I., et al. (2024). GenCast: diffusion-based ensemble forecasting for medium-range weather. Nature, 637, 1038–1044.
9. Sønderby, C.K., et al. (2020). MetNet: a neural weather model for precipitation forecasting. arXiv preprint arXiv:2003.12140.
10. Ronneberger, O., et al. (2015). U-Net: convolutional networks for biomedical image segmentation. MICCAI 2015.
11. Akiba, T., et al. (2019). Optuna: a next-generation hyperparameter optimization framework. KDD 2019.
12. SHMI (2024). Annual report on meteorological measurements in Slovakia 2024. Bratislava: SHMI.