Outperforming Self-Attention Mechanisms in Solar Irradiance Forecasting via Physics-Guided Neural Networks
Home / Papers / Outperforming Self-Attention Mechanisms in Solar Irradiance Forecasting via Physics-Guided Neural Networks
Outperforming Self-Attention Mechanisms in Solar Irradiance Forecasting via Physics-Guided Neural Networks
Authors: Mohammed Ezzaldin Babiker Abdullah, R. R Mohammed Date: 2026-04-15 Paper ID: openalex:2604.13455
Summary
This paper challenges the trend of using complex, computation-heavy Transformer architectures for solar irradiance forecasting by introducing a physics-informed hybrid CNN-BiLSTM model. By incorporating domain-specific physical constraints and engineered features like the Solar Zenith Angle, the proposed framework achieves superior performance in high-noise environments compared to self-attention mechanisms. The authors demonstrate that integrating domain knowledge is more effective than architectural scaling for real-time renewable energy management, providing a significant reduction in RMSE on NASA POWER data.
Key Contributions
- Introduces a physics-informed hybrid CNN-BiLSTM architecture that integrates 15 domain-specific features (e.g., Clear-Sky indices) to outperform Transformer-based models in GHI forecasting.
- Demonstrates the ‘Complexity Paradox’ in meteorological forecasting, showing that explicit physical constraints provide superior accuracy (RMSE 19.53 W/m^2) and efficiency compared to high-complexity self-attention mechanisms.
- Validates the proposed approach using NASA POWER solar radiation datasets in Sudan, establishing a new baseline for high-noise, arid-region irradiance forecasting.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.13455
- paper_id
- 2604.13455
- paper_source
- openalex
- domain
- time-series
- tags
- time-series-forecastingphysics-informed-mlneural-networks
- architectures
-
- datasets
-
- concept_slugs
-
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-18T05:34:53Z