A Hybrid Framework for Long-Term Photovoltaic Power Generation Forecasting

Recent extreme weather events around the world have increased interest in renewable energy. As Photovoltaic power grows in importance, so does the need for long-term forecasting. Generally, the data related to Photovoltaic power generation consists of many variables with multi-periodicity. In this study, we propose a new architecture that combines TimesNet and iTransformer models to maximize the performance of long-term Photovoltaic power generation forecasting. First, the TimesNet model is used to identify intra- and inter-period variations and to transform 1D information into 2D to effectively model different time series patterns in temporal data. Second, the iTransformer model is utilized to capture multivariate correlations using the Attention mechanism to reflect relationships across the entire time series. The proposed model outperformed the existing models in Photovoltaic power generation forecasting.