Ultra-Short-Term Forecasting of Photovoltaic Power Generation through Spatiotemporal Time-Series Image Conversion

Abstract

Seasonal fluctuations and the intermittent nature of photovoltaic (PV) generation create significant challenges for accurate short-term forecasting. This study presents Next Frame Gramian Angular field U-Net (NFGUN), a hybrid deep learning forecasting framework that stands apart from conventional methods by transforming 1D PV time-series data into 2D Gramian Angular Summation Field (GASF) images. Unlike models that rely on direct regression or sky imagery, NFGUN forecasts the next GASF frame using a deep architecture and reconstructs it back into time-series form, effectively capturing nonlinear temporal dynamics. Its uniqueness lies in several key innovations: (1) the integration of Convolutional Long Short-Term Memory 2D (ConvLSTM2D) into a customised U-Net model for better generalisation spatiotemporal features; (2) the incorporation of residual blocks in the bottleneck to preserve deep features while mitigating vanishing gradients and cyclical encoding of time to enrich seasonal patterns; (3) the use of Lanczos interpolation with CIEDE2000 colour difference for high-precision reconstruction from predicted image frames. We evaluate NFGUN against six well-established forecasting methods and measure performance using six accuracy metrics such as MAE, RMSE, and WAPE across all four seasons; NFGUN demonstrates superior performance. Compared to the best-performing benchmark, it achieved improvements in MAE (61.23% winter, 56% spring, 37.45% summer, 59.67% autumn), RMSE (48.34% winter, 64.63% spring, 31.65% summer, 45.83% autumn), and WAPE (49.9% winter, 43.84% spring, 45.83% summer, 48.72% autumn), underscoring its ability to adapt to seasonal variability. These results demonstrate NFGUN's ability to effectively capture complex, seasonal dynamics, making it a robust solution for ultra-short-term PV power forecasting.