Transformer-based temporal models for probabilistic load and photovoltaic power forecasting in commercial microgrids

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Renewable Energy Systems

Background:

• Effective energy management in commercial microgrids requires accurate forecasting of solar photovoltaic (PV) generation and load demands, accounting for inherent uncertainties.
• Existing forecasting models exhibit instability and reduced accuracy (8-15%) when forecasting horizons exceed 12 time steps, limiting their scalability and reliability.
• Limited research addresses the impact of rolling-horizon forecasting on managing uncertainties in smart grid environments.

Purpose of the Study:

• To develop and validate a transformer-based forecasting framework for solar PV and load in smart grid applications.
• To improve probabilistic forecasting accuracy and stability, particularly for longer forecasting horizons.
• To enhance the estimation of uncertainty through horizon-aware learning mechanisms.

Main Methods:

• Development of a novel transformer-based PV-load forecasting framework incorporating horizon-aware learning mechanisms.
• Implementation and simulation of the model in a MATLAB environment using rolling-horizon experiments.
• Evaluation of the model's performance using Continuous Ranked Probability Score (CRPS) and pinball loss under three different operational conditions.

Main Results:

• Achieved probabilistic forecasting accuracy greater than 12%, with a 12.6% reduction in CRPS.
• Demonstrated effective prediction interval capture with 9.4% internal coverage and a 18% reduction in computational cost.
• The transformer-based model outperformed Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) models, showing 12% improvement in average and 6% in worst-case scenarios.

Conclusions:

• The proposed transformer-based framework offers superior performance in probabilistic PV-load forecasting compared to existing techniques.
• The model exhibits scalability up to 14 forecasting horizons with consistent stability and is suitable for low-latency applications (<1.5s).
• Future work can explore extending the transformer-based concept beyond 14 horizon steps to further analyze performance degradation.