Photovoltaic Generation Prediction using a Weather-Forecast-Aware Autoformer

Effective management of Distributed Energy Resources is essential for a green-energy transition, ensuring sustainable and reliable power. This requires accurately forecasting complex renewable energy production, a task well-suited for deep neural networks due to their ability to model intricate temporal and spatial dynamics with low computational overhead compared to physics-based simulation. This work leverages a domain-specific normalization strategy and a modified Autoformer, a Transformer-like neural network architecture that uses the Cross-Correlation function for calculating time dependencies, to better capture signal periodicities and leverage future information. The resulting scheme was benchmarked against established forecasting architectures using real-world photovoltaic generation data, resulting in comparable accuracy with the added benefit of interpretable internal coefficients, which are shown to be a useful analytical tool.