FE-ConvGRU: A novel deep learning framework for ultra-short-term PV power prediction by encoding temporal geometry and attention mechanism

Amidst the global imperative for renewable energy, advanced ultra-short-term photovoltaic (PV) power prediction technology is urgently needed for grid stability in new hybrid power systems. Existing methods primarily rely on exogenous variables to enhance power variation features. Yet such variables are not always satisfied due to sensor failures or transmission delays in real testing applications, and models that use combination variables as inputs may not function, which raises a higher demand for models to mine intrinsic signatures of power sequences far more rigorously. Therefore, a novel deep learning framework, called the feature-enhanced convolutional gated recurrent unit (FE-ConvGRU) network, is designed in this study to further improve prediction performance under extreme scenarios where solely PV power sequences are available. Firstly, the Gramian angular field (GAF) transformation is introduced to encode temporal power sequences as geometric features, enhancing the capture of periodic characteristics and power transients. Secondly, a coordinate attention (CA) mechanism is embedded within convolutional layers to perform multi-channel spatial feature fusion, preserving long-range dependencies and directional patterns across encoded geometric features. Thirdly, the GRU-based prediction module processes those enhanced features for sequential predictions under diverse multiple time horizons. Finally, extensive experiments on two different scaled PV power systems demonstrated the superiority of the proposed method, whether across different PV power stations or under diverse prediction horizons, achieving the lowest prediction errors compared to comparison methods.