Solar power forecasting based on domain adaptive learning

Abstract

Solar power forecasting is critical to ensure the safety and stability of the power grid with high photovoltaic power penetration. Machine learning methods are compelling in solar forecasting. These methods can capture the complex coupling relationship between different meteorological factors without physical modeling. Most of the existing machine learning based forecasts follow the batch learning manner. Once the training is completed, the structure and parameters of the model are usually no longer adjusted. However, the climate is complex and dynamic. It is difficult for a fixed model to adapt to the climate characteristics of different regions or periods. Therefore, an online domain adaptive learning approach is proposed in this paper. Knowledge can be selectively accumulated or forgotten in its iterative process. As weather changes, the model can dynamically adjust its structure to adapt to the latest weather conditions. Unlike existing adaptive iterative methods, the proposed adaptive learning approach does not rely on the labels of the test data in the updating process. Experiments show that this method can effectively track changes in data distribution and obtain reliable prediction results.