A hybrid framework for short-term solar power prediction using signal decomposition and attention-based network

The integration of solar power into modern electrical grids is a critical step toward achieving sustainable energy goals. However, the inherent variability and non-stationarity of solar power generation pose significant challenges for accurate forecasting. This paper presents an enhanced hybrid forecasting model that combines the signal decomposition power of Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) with a novel deep learning architecture featuring a temporal convolutional network (TCN), a gated recurrent unit (GRU), and a self-attention mechanism. The model was developed and rigorously evaluated using a comprehensive, real-world dataset collected from a rooftop photovoltaic system in Tay Ninh province, Vietnam. The proposed CEEMDAN-TCN-GRU-Attention model demonstrates substantial improvements over established forecasting benchmarks, including eXtreme Gradient Boosting, a standalone long short-term memory, and a CEEMDAN-GRU hybrid counterpart. Notably, the model achieves an impressive R2 value of 98.24%, underscoring its superior ability to capture the complex dynamics of solar power generation. The model's success stems from the synergistic combination of robust signal decomposition to handle non-stationarity and an advanced sequential architecture that effectively extracts features, models temporal dependencies, and focuses on salient information. This research contributes to enhancing grid stability by providing a more reliable and accurate tool for short-term solar power forecasting.