Brum et al. (2026) Assessing the environmental costs of multi-scale recurrent neural networks for sustainable extreme rainfall nowcasting

Identification

• Journal: Scientific Reports
• Year: 2026
• Date: 2026-03-10
• Authors: Douglas Brum, Luan Teylo, Fabrício Polifke da Silva, Fernanda Cerqueira Vasconcellos, Gabriel Berto Breder, Lívia de Azevedo, Eduardo Bezerra, Fábio Porto, Mariza Ferro
• DOI: 10.1038/s41598-026-43029-2

Research Groups

• Fluminense Federal University, Institute of Computing, Niterói-RJ, Brazil
• Centre Inria de l’université de Bordeaux, Talence, France
• Universidade Federal do Rio de Janeiro, Meteorology Department, Rio de Janeiro, Brazil
• Federal Center for Technological Education of Rio de Janeiro, Computer Science Department, Rio de Janeiro, Brazil
• National Laboratory of Scientific Computation, Petrópolis, Brazil

Short Summary

This study evaluates the Multi-scale Recurrent Neural Network (MS-RNN) framework for improving computational efficiency and predictive accuracy in extreme precipitation nowcasting using real weather radar data. It quantifies the environmental costs (energy, CO2 equivalent emissions, and water usage) of deep learning models to support sustainable and accessible AI solutions for climate resilience in resource-limited regions.

Objective

• Evaluate the effectiveness of the Multi-scale Recurrent Neural Network (MS-RNN) framework for improving computational efficiency and predictive accuracy in extreme precipitation nowcasting.
• Quantify the environmental costs (energy consumption, CO2 equivalent emissions, and water usage) of deep learning models for precipitation nowcasting, especially when enhanced by the MS-RNN framework, to assess the feasibility of sustainable AI for extreme rainfall prediction.

Study Configuration

• Spatial Scale:
  • TAASRAD19 dataset: 240 km diameter, 500 m resolution, images 480x480 pixels.
  • Rio de Janeiro dataset: 250 km radius, approximately 500 m resolution, projected onto 17x29 points, zero-padded to 32x32 pixels, covering the metropolitan region of Rio de Janeiro.
• Temporal Scale:
  • TAASRAD19 dataset: 300 second temporal resolution. Data from 2010-2013 (training 2010-2011, validation 2012, testing 2013).
  • Rio de Janeiro dataset: 600 second temporal resolution. Data from 2016-2022 (training 2016-2018, validation 2019-2020, testing through April 2021-2022).
  • Nowcasting lead times: 0-7200 seconds.

Methodology and Data

• Models used: PredRNN-V2, MotionRNN, ConvLSTM, and their MS-RNN-enhanced counterparts (MS-PredRNN-V2, MS-MotionRNN, MS-ConvLSTM).
• Data sources:
  • TAASRAD19 dataset: Weather radar reflectivity data from the Italian Alps (2010-2013).
  • Rio de Janeiro dataset: Weather radar reflectivity data from the State Environmental Institute (INEA) in Rio de Janeiro, Brazil (2016-2022).
  • Environmental metrics: nvidia-smi (GPU energy consumption), CodeCarbon library (CO2 equivalent emissions, water usage).

Main Results

• Computational Efficiency:
  • MS-RNN reduced GPU memory usage by 41-50% on TAASRAD19 (e.g., average reduction of 16.4 x 10^9 bytes to 20 x 10^9 bytes from a 40 x 10^9 byte baseline) and 22.3-24.3% on Rio de Janeiro.
  • MS-RNN reduced execution time by 50-66% on TAASRAD19 (e.g., average reduction of 64,800 seconds from a 126,000 second baseline for ConvLSTM) and 6.81-10.40% on Rio de Janeiro.
• Environmental Impact:
  • MS-RNN reduced GPU energy consumption by 54.45-71.05% on TAASRAD19 (e.g., average reduction of 18 MJ from a 27 MJ baseline for ConvLSTM) and 22.40-24.50% on Rio de Janeiro.
  • MS-RNN reduced CO2 equivalent emissions by 31.3-61.7% on TAASRAD19 (e.g., from 0.60 kg to 0.23 kg for PredRNN-V2) and 14.3-15.4% on Rio de Janeiro.
  • MS-RNN reduced water consumption by 31.3-62.7% on TAASRAD19 (e.g., from 39.54 x 10⁻³ m³ to 14.76 x 10⁻³ m³ for PredRNN-V2) and 9.4-11.9% on Rio de Janeiro.
  • Total project environmental cost: approximately 1754.28 MJ, 47.01 kg CO2 equivalent, and 8.46 m³ water.
• Predictive Accuracy:
  • On TAASRAD19, MS-RNN-enhanced models maintained or slightly improved predictive performance (e.g., lower Mean Squared Error (MSE) and Mean Absolute Error (MAE), similar Structural Similarity Index (SSIM), improved Critical Success Index (CSI) at higher thresholds).
  • On Rio de Janeiro, MS-RNN-enhanced models showed a slight reduction in predictive accuracy (e.g., SSIM decreased by up to 3.3%, MSE increased by up to 7.2%) compared to baselines, but still delivered competitive results.
  • MS-enhanced models demonstrated superior accuracy at higher rainfall intensity thresholds (2.778 x 10⁻⁶ m/s (10 mm/h) and 8.333 x 10⁻⁶ m/s (30 mm/h)) for CSI scores on TAASRAD19.
  • All models, including MS-RNN variants, still struggle to forecast heavy rainfall events (evidenced by high balanced MSE and MAE).
  • Extreme precipitation is defined as a rainfall intensity equal to or greater than 1.389 x 10⁻⁵ m/s (50 mm/h), but CSI was reported up to ≥8.333 x 10⁻⁶ m/s (30 mm/h) due to data rarity.

Contributions

• First study to quantify the environmental costs (energy consumption, CO2 equivalent emissions, and water usage) of deep learning models for precipitation nowcasting, particularly when enhanced by the MS-RNN framework.
• Demonstrates that the MS-RNN framework significantly reduces computational and environmental costs while maintaining competitive predictive accuracy, supporting sustainable AI for climate resilience in resource-limited regions.
• Highlights the critical need to balance predictive precision with sustainability in climate-related AI applications, especially in the Global South.
• A systematic literature review revealed a critical gap: no existing studies apply Green AI principles to precipitation nowcasting.

Funding

• Serrapilheira Institute (grant number Serra – 2211-41897)
• SERRAPILHEIRA/FAPERJ – 2023 (Proc. E-26/210.242/2024)

Citation

@article{Brum2026Assessing,
  author = {Brum, Douglas and Teylo, Luan and Silva, Fabrício Polifke da and Vasconcellos, Fernanda Cerqueira and Breder, Gabriel Berto and Azevedo, Lívia de and Bezerra, Eduardo and Porto, Fábio and Ferro, Mariza},
  title = {Assessing the environmental costs of multi-scale recurrent neural networks for sustainable extreme rainfall nowcasting},
  journal = {Scientific Reports},
  year = {2026},
  doi = {10.1038/s41598-026-43029-2},
  url = {https://doi.org/10.1038/s41598-026-43029-2}
}