Trend-Context Fusion Network with Multi-Head Attention for Solar Photovoltaic Power Forecasting

Accurate forecasting of photovoltaic (PV) power generation is essential for maintaining smart grid stability and supporting efficient renewable energy management. This study presents a hybrid deep learning framework that integrates one-dimensional convolutional neural networks (1D-CNNs), multi-head attention mechanisms, and long short-term memory (LSTM) networks to enhance short-term PV power prediction performance. Within the proposed architecture, the 1D-CNN component captures localized high-frequency temporal patterns, while the multi-head attention module models long-range contextual dependencies across time. The LSTM network subsequently learns the sequential dynamics governing PV power generation. To improve robustness and generalization, cyclical time-based features reflecting diurnal and seasonal characteristics are incorporated, and a systematic hyperparameter optimization strategy is employed. The proposed model is evaluated using real-world meteorological and PV power generation datasets collected from Gwangyang Port and the Dangjin Landfill Solar Power Plant in South Korea. Experimental results demonstrate that the proposed hybrid framework consistently outperforms conventional single-model baselines and alternative deep learning architectures. Notably, it achieves lower mean absolute error (MAE) and root mean square error (RMSE), along with a higher coefficient of determination (R²), indicating superior predictive accuracy and explanatory capability. Furthermore, model interpretability is enhanced through SHapley Additive exPlanations (SHAP), which quantify the relative contributions of key meteorological variables to PV power output. This interpretability analysis provides transparent insights into model behavior and supports informed decision-making for advanced energy management and smart grid operation.