An attention-based Bayesian sequence to sequence model for short-term solar power generation prediction within decomposition-ensemble strategy

The utilization of renewable energy has attracted much attention with the deterioration of the environment. Solar energy is widely distributed and has huge reserves. But, solar energy is also volatile and intermittent. This brings difficulties to the large-scale grid connection of solar energy and the distribution of power resources. The key to solve this problem is to provide accurate and stable short-term prediction for solar power generation. For this purpose, we present an attention-based Bayesian sequence to sequence (Seq2Seq) model within decomposition-ensemble strategy. Firstly, an influential factors analysis is performed on the data about meteorological conditions, operation status of the photovoltaic panels and historical solar power generation sequences to obtain the optimal combination of the influential factors. Secondly, this paper proposes a novel decomposition-ensemble framework based on complete ensemble empirical mode decomposition with adaptive noise and independent component analysis to mine the intrinsic modes of the solar power generation time series. Thirdly, this paper presents an attention-based Bayesian Seq2Seq method for modeling the relationships between solar power generation and influential factors. Using a real-world dataset from State Administration of Electricity Investment Science and Technology of China, a case study, comparative analysis and robustness checks are conducted, through which the contributions of the methods in the proposed model and the superior performance of the model are demonstrated.