Enhancing Solar Energy Production Forecasting Using Advanced Machine Learning and Deep Learning Techniques: A Comprehensive Study on the Impact of Meteorological Data

This study explores advanced machine learning (ML) and deep learning (DL) techniques for improving the accuracy of solar energy production forecasting. By integrating detailed meteorological data and novel modelling approaches, the research demonstrates the impact of data preprocessing, feature selection, and hybrid models on prediction performance. The best-performing approach combined LSTM with Inception layers, leveraging a rule-based prediction wrapper and a cascaded stacking model to refine forecasting accuracy.

Takeaways:

1. The study highlights the importance of meteorological data in solar energy forecasting, showing how environmental factors influence photovoltaic output.
2. Machine learning models, including SVM, KNN, and CatBoost, were tested alongside deep learning approaches, such as LSTM and LSTM with Inception layers.
3. A rule-based prediction wrapper was introduced to restrict predictions to daylight hours, improving model performance by integrating sunrise and sunset data.
4. A cascaded stacking model was proposed, combining weak predictors, association rules, and an enhanced aggregation procedure to improve generalization.
5. Feature selection and preprocessing, including logarithmic transformations and encoding techniques, significantly impacted model accuracy.
6. The LSTM with Inception layers model outperformed other models, demonstrating the potential of hybrid deep learning techniques for solar energy forecasting.
7. Ensemble learning methods, such as gradient boosting and random forests, helped reduce prediction errors by integrating multiple base models.
8. The study provides insights into real-world applications, particularly for Ukraine’s energy security, emphasizing the need for distributed generation and robust forecasting models.