Ultra-Short Term Photovoltaic Generation Forecasting Based on Data Decomposition and Customized Hybrid Model Architecture

Abstract

As photovoltaic (PV) systems have been successfully adopted worldwide, accurate power generation forecasting becomes increasingly essential to stable power grid operation and smart grid applications to cope with the variability of PV systems. Several data-driven models have recently been proposed for the more accurate prediction of PV power generation and have shown good performance. In particular, hybrid models that combine the characteristics of single-structure deep learning-based models have achieved better accuracies. To this end, a novel ultra-short term PV power generation forecasting model with a hybrid structure is proposed for instantaneous response to PV fluctuations. For higher forecasting accuracy, the proposed model decomposes the input feature data into trend and residual components and employs customized sub-models such as the linear, Transformer, and long short-term memory (LSTM). Furthermore, the proposed model is trained with data from the self-built PV site to implement a model suitable to real-world applications. Finally, the experimental results demonstrate that the proposed model has the best forecasting performance compared to conventional and state-of-the-art deep learning-based forecasting models with reasonable computational complexity.