Intra-Day Solar Power Forecasting Strategy for Managing Virtual Power Plants.

Area of Science:

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

Background:

• Rising solar energy penetration necessitates accurate short-term power forecasting for stable grid operation.
• Inaccurate forecasts hinder the integration of solar resources and discourage their use.
• Virtual Power Plants (VPPs) offer a centralized management solution to minimize forecasting errors.

Purpose of the Study:

• To introduce an efficient method for accurate intra-day Photovoltaic (PV) power forecasts at multiple locations using readily available data.
• To incorporate prediction intervals based on Mean Absolute Error (MAE) to quantify forecast uncertainty for VPP node power generation.

Main Methods:

• Utilized a Long Short-Term Memory (LSTM) neural network for PV power forecasting.
• Emulated a VPP using a real PV installation and proximate ground-based meteorological stations.
• Calculated Mean Absolute Error (MAE) to evaluate forecast accuracy and uncertainty.

Main Results:

• Achieved a Mean Absolute Error (MAE) of 44.19 W/m² for PV power forecasts, comparable to other deep learning methods.
• Demonstrated a 12.37% reduction in global error (MAE) when applying the technique across 8 VPP nodes.
• Prediction intervals provided valuable insights into VPP node power generation uncertainty.

Conclusions:

• The proposed LSTM-based forecasting method is effective for producing accurate intra-day PV power predictions.
• Integrating this forecasting technique within VPPs significantly reduces overall forecasting errors.
• The approach shows substantial potential for enhancing the management and reliability of distributed solar energy resources.