Effect of Climate on Photovoltaic Yield Prediction Using Machine Learning Models

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence in Energy
• Climate Science and Energy

Background:

• Machine learning (ML) is increasingly applied to photovoltaic (PV) power prediction.
• The influence of diverse climatic conditions on ML-based PV yield predictions remains largely unexplored.
• A comprehensive dataset of PV systems across various climates was compiled and made publicly available.

Purpose of the Study:

• To investigate the impact of climate on the accuracy of ML models for PV power prediction.
• To identify climatic trends affecting PV energy yield forecasting.
• To establish a benchmark for ML model performance across different climate zones.

Main Methods:

• Trained five ML algorithms and a baseline model on data from 48 PV systems globally, divided equally into four climate types.
• Prioritized open-data sources for data gathering, creating a dedicated website for accessibility.
• Conducted robustness evaluations to ensure the reliability of the findings.

Main Results:

• Algorithm performance rankings were consistent across all climates.
• Dry climates exhibited the lowest average Normalized Root Mean Squared Error (NRMSE) at 47.6%, while tropical climates showed the highest at 60.2%.
• Models trained in cold climates demonstrated better generalization to other climates, with an average additional NRMSE of 5.6%.

Conclusions:

• Climate significantly influences the accuracy of ML-based PV power predictions, with dry climates being most favorable.
• The generalization capability of ML models varies by climate, with cold-climate-trained models showing superior cross-climate predictive performance.
• The study provides valuable insights and open-access data for advancing ML applications in PV energy forecasting under diverse climatic conditions.