MATNet: multi-level fusion transformer-based model for day-ahead PV generation forecasting

Area of Science:

• Electrical Engineering
• Computer Science
• Renewable Energy Systems

Background:

• Accurate forecasting of renewable energy generation, particularly from photovoltaic (PV) systems, is critical for grid stability and efficient integration of renewable energy sources (RESs).
• Current Artificial Intelligence (AI)-based forecasting models excel at pattern recognition but often underutilize physics-derived information from Numerical Weather Prediction (NWP) models.

Purpose of the Study:

• To propose and evaluate MATNet, a novel transformer-based multimodal architecture for multi-step day-ahead PV power generation forecasting.
• To enhance PV forecasting accuracy by effectively fusing historical PV data with NWP-derived weather covariates.

Main Methods:

• Development of MATNet, a transformer-based multimodal architecture utilizing a multi-level joint fusion approach with soft-attention mechanisms.
• Input data includes historical PV generation data and historical/future weather covariates from NWP models.
• Evaluation on the Ausgrid benchmark dataset and five external PV datasets, including ablation studies, sensitivity analysis, and cross-site zero-shot evaluation.

Main Results:

• MATNet achieved a Root Mean Square Error (RMSE) of 0.0445, representing a 65% relative improvement over the best-performing baseline.
• Experiments demonstrated MATNet's resilience to missing data, strong transferability across different PV sites, and a favorable accuracy-computational cost trade-off.
• The multi-level joint fusion approach effectively leverages both PV and weather data.

Conclusions:

• MATNet offers a reliable and efficient solution for multi-step day-ahead PV power forecasting.
• The proposed architecture facilitates better integration of PV energy into the power grid by improving forecasting accuracy and robustness.
• The study highlights the benefit of combining AI with physics-informed data for renewable energy forecasting.