A Comparative Study of Power Forecasting Models for Renewable Energy: A Case Study of the Azores, Portugal

Power forecasting in off-grid microgrids presents distinct difficulties, primarily arising from the systems’ geographic isolation and the integration of diverse energy sources. This paper provides a comprehensive evaluation of state-of-the-art time-series forecasting models for power prediction in off-grid microgrids.Using real-world data from an isolated microgrid located on one of the Azores Islands-where wind, solar, and diesel energy sources are jointly operated. The study compares best-in-class models based on Transformer, recurrent neural network, convolutional neural network, and linear architectures. The experimental results show that PatchTST and PatchMixer excel in capturing local dependencies and nonlinear features compared to more complex models, while simpler linear models such as DLinear and NLinear also demonstrate competitive performance. The findings indicate that patch-based input representations and linear structures are particularly effective for off-grid renewable energy datasets and are directly applicable to islanded and other off-grid environments, providing clear insights into the strengths and limitations of different forecasting approaches.