Spectral-temporal convolutional approach for PV systems output power forecasting: Case studies in single-site and multi-site

Abstract

Accurate predictions of photovoltaic (PV) and wind power outputs are indispensable for integrating additional renewable energy sources into the grid. Photovoltaic energy is influenced by weather conditions and cloud movement, constituting the primary factors contributing to stochastic characteristics and uncertainty in predictive accuracy. Consequently, recent studies have focused on developing forecasting methods that utilize weather factors or nearby farms as multi-variable inputs. This study introduces several challenges, and one of the most significant goals is to capture the correlation between data series (spatial feature) and temporal correlation within each series (temporal feature) simultaneously. However, most previous work has concentrated solely on the time or frequency domain. This work presents a novel approach based on the spectral and frequency domain, which efficiently captures spatial–temporal correlations via Graph Fourier Transform (GFT) and Discrete Fourier Transform (DFT). In this manner, spatial correlations are clearly represented via the adjacency matrix of GFT, and temporal correlations are clarified via DFT. The study provides specific evidence concerning the interpretability and extraction efficiency of the adjacency matrix in the context of multi-variable, single-site, and multi-site data. Simultaneously, it demonstrates the effectiveness of the representation approach in the spectral domain. The results pointed out the forecasting method proposed has higher forecasting accuracy than the benchmark model.