Improved day-ahead PV power forecasting through global multi-dimensional coordinate attention and feature fusion techniques

High penetration of photovoltaic (PV) generation introduces intermittency and uncertainty, challenging the stability and dispatch of modern power systems. Accurate PV power forecasting is crucial for mitigating these challenges. While existing hybrid deep learning models have advanced spatiotemporal feature extraction, they still struggle with modeling long-range dependencies and cross-hourly irradiance-power dynamic couplings, which has become a critical bottleneck limiting long-term forecasting accuracy.To address this issue, this paper proposes an innovative Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN)-enabled Global Attentive Fusion Network (CGAFN) for refined PV power forecasting. The CGAFN leverages CEEMDAN to decompose and capture both the regular and stochastic components of PV power. It introduces a novel global multi-dimensional coordinate attention module specifically designed to capture spatial heterogeneity and channel correlations in multi-source meteorological data for fine-grained feature extraction, and models global temporal dependencies using a hybrid Long Short-Term Memory (LSTM)-Transformer network. Experimental results on a real-world 20 MW PV plant dataset demonstrate that the proposed model achieves superior performance in day-ahead multi-step PV power forecasting, with average mean absolute error (MAE), mean squared error (MSE), and coefficient of determination (R²) reaching 1.1677 MW, 3.1143 MW², and 0.8843, respectively. Compared with the mainstream LSTM benchmark, it achieves significant accuracy improvements: MAE and MSE decrease by 22.7% and 38.7%, respectively, and R² increases by 9.0%.Ablation analysis further demonstrates the effectiveness of each core module and their synergistic effects. The hybrid model provides methodological reference for multi-source time series data-driven energy system forecasting tasks.