SolarTformer: A Transformer Based Deep Learning Approach for Short Term Solar Power Forecasting
SolarTformer: A Transformer Based Deep Learning Approach for Short Term Solar Power Forecasting
Authors: Ankan Basu, J.C. Roy, Aditya Datta, Prayas Sanyal, Sumanta Banerjee Date: 2026-04-27 Paper ID: openalex:2604.24306
Summary
This paper introduces SolarTformer, an attention-based deep learning model designed for short-term solar power forecasting using meteorological inputs. By utilizing self-attention mechanisms, the architecture effectively captures intricate temporal dependencies and spatial variability in solar irradiance. Furthermore, the inclusion of power station-specific metadata enhances the model’s ability to generalize across different site configurations and seasonal changes, outperforming existing forecasting baselines.
Key Contributions
- Introduces SolarTformer, a transformer-based architecture for short-term solar power forecasting leveraging self-attention for complex temporal and spatial dependencies.
- Integrates station-specific metadata into the forecasting pipeline to improve generalization across diverse geographical locations, panel configurations, and seasonal variations.
- Demonstrates significant performance improvements over state-of-the-art models, maintaining high robustness during both clear and cloudy weather conditions.
Open Questions & Future Work
Archivist Review
I have rejected the model name ‘SolarTformer’ as it is a specific implementation instance. The open questions regarding edge efficiency and geographic generalization are highly relevant to the broader field of solar energy forecasting and deserve tracking as industry-wide challenges.
Approved Open Questions
- Efficiency on Edge Devices: This is a critical bottleneck for real-world adoption of high-accuracy deep learning models in remote PV stations where hardware resources are limited.
- Geographic and Climatic Generalization: Generalization is the primary challenge in scaling solar forecasting solutions to global power grids, where weather patterns and PV system configurations vary significantly.
Rejected Candidates
- [concept] SolarTformer (
solartformer) - paper_local: This is a paper-specific model name rather than a reusable architectural concept.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.24306
- paper_id
- 2604.24306
- paper_source
- openalex
- domain
- time-series
- tags
- time-series-forecastingtransformerenergy-management
- architectures
-
- datasets
-
- concept_slugs
-
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-30T06:03:21Z