Forecasting PV Energy Generation Using Transformer-Based Architectures: A Comparative Study of Lag-Llama, TFT, and DeepAR

This paper investigates the effectiveness of transformer-based models, the Temporal Fusion Transformer (TFT) and Lag-Llama, alongside the recurrent neural network-based model DeepAR, in forecasting photovoltaic (PV) energy generation. With the rising integration of renewable energy sources, precise forecasting has become critical for energy grid management, storage system operation, and maintaining grid stability. The study evaluates the performance of these models across different seasonal test sets, employing metrics such as Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). Our results indicate that while no single model consistently outperforms the others across all seasons, the fine-tuned Lag-Llama model demonstrates superior accuracy in most conditions, except during the summer, where DeepAR exhibits better performance. This paper highlights the potential of using advanced transformer-based models for time series forecasting in the renewable energy sector, suggesting that hybrid approaches may enhance future forecasting accuracy.