Solar Power Forecasting Using Hybrid Deep Learning: Performance Enhancement with Random Forest-BiLSTM and Ensemble...

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence
• Time Series Analysis

Background:

• Accurate solar power forecasting is crucial for the stability and grid integration of renewable energy systems.
• Deep learning models offer potential for capturing complex temporal dependencies in solar irradiance data.

Purpose of the Study:

• To develop and evaluate a hybrid deep learning ensemble approach for solar generation forecasting.
• To assess the performance of various hybrid architectures, including RF-BiLSTM, CNN-LSTM, CNN-BiLSTM, CNN-GRU, and CNN-Transformer.

Main Methods:

• Five hybrid deep learning architectures were implemented, combining convolutional and recurrent neural networks.
• Models were trained and evaluated on historical solar irradiance time series data.
• An ensemble model was created using inverse Mean Absolute Error (MAE) weighted averaging of the top three performing individual models.

Main Results:

• The RF-BiLSTM model showed the best individual performance (R² = 0.6568, MAE = 30,728 W).
• The ensemble model achieved superior results (R² = 0.6933, MAE = 28,809.89 W), reducing prediction error by 6.2%.
• The ensemble approach demonstrated enhanced forecast robustness and accuracy.

Conclusions:

• The proposed hybrid deep learning ensemble framework effectively improves solar power forecasting accuracy.
• This data-driven solution offers a scalable and robust method for renewable energy forecasting in smart grids.
• The findings contribute to better grid integration and operational stability for solar energy systems.