Rethinking the use of deep learning methods for photovoltaic power forecasting.

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence in Energy
• Time-Series Forecasting

Background:

• Accurate photovoltaic (PV) power forecasting is essential for stable grid operations.
• Challenges include weather variability and integrating historical data with future predictions.
• Deep learning models require optimized architectures for time-series PV forecasting.

Purpose of the Study:

• To re-evaluate deep learning architectures for PV power forecasting.
• To demonstrate the benefits of full encoder-decoder models and channel dependence.
• To propose an improved forecasting model integrating historical data and weather forecasts.

Main Methods:

• Proposed Cross-Unet, a Transformer-based architecture with multi-scale temporal encoding.
• Implemented correlation-aware channel attention and hierarchical cross-attention decoding.
• Utilized numerical weather prediction, satellite irradiance, and AI weather model forecasts.

Main Results:

• Cross-Unet outperformed ten deep learning baselines and traditional benchmarks.
• Achieved superior performance across diverse PV power stations and forecasting horizons (4 hours to 7 days).
• Demonstrated effectiveness with various forecast sources, including AI-based weather models.

Conclusions:

• Cross-Unet effectively fuses historical generation data with weather forecasts.
• Enables operational 15-minute-resolution PV power predictions for grid scheduling and energy trading.
• Highlights the importance of advanced architectures and AI weather models for accurate solar energy forecasting.