Transfer learning strategies for solar power forecasting under data scarcity.

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence in Energy
• Smart Grid Technologies

Background:

• Accurate solar power forecasting is crucial for grid stability and energy management in smart cities.
• Limited data from new or unmetered solar plants hinders traditional forecasting model training.

Purpose of the Study:

• To evaluate the effectiveness of transfer learning (TL) strategies for improving solar plant production forecasts.
• To address forecasting challenges posed by insufficient data in solar energy systems.

Main Methods:

• Utilized a stacked Long Short-Term Memory (LSTM) model incorporating three transfer learning strategies.
• Applied TL for both weight initialization and feature extraction with varying freezing techniques.
• Compared TL models against conventional non-TL and smart persistence models.

Main Results:

• TL models demonstrated significant performance improvements over conventional methods.
• Achieved 12.6% higher accuracy (RMSE) and 16.3% better forecast skill index with one year of data.
• The performance gap widened considerably with only three months of training data.

Conclusions:

• Transfer learning is a reliable and effective tool for accurate solar production forecasting, particularly for new installations.
• TL enables better energy balancing and demand response management, even with limited historical data.
• This research advances power production forecasting for self-producers and energy communities.