Explaining solar forecasts with generative AI: A two-stage framework combining transformers and LLMs.

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence
• Machine Learning

Background:

• Accurate solar power forecasting is essential for integrating Photo-Voltaic (PV) systems into the energy grid.
• Current forecasting methods often lack interpretability, hindering trust and adoption.

Purpose of the Study:

• To develop a novel two-stage hybrid framework for accurate and interpretable solar power forecasting.
• To enhance the integration of PV systems into modern energy infrastructure.

Main Methods:

• A Transformer-based encoder-decoder architecture (SolarTrans) was used for short-term DC power prediction.
• Multivariate time series data including weather and inverter data were utilized.
• A generative Large Language Model (LLM), Flan-T5, was fine-tuned for generating natural language explanations of forecasts.

Main Results:

• The SolarTrans model demonstrated strong predictive performance on PV plant datasets, achieving low Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE), and high R2 scores.
• The explanation module generated high-fidelity, domain-relevant natural language explanations, evidenced by strong ROUGE and BLEU scores.

Conclusions:

• The proposed hybrid framework effectively enhances solar power forecast accuracy and interpretability.
• This approach facilitates better integration of PV systems by providing reliable and understandable predictions.