Short-term distributed photovoltaic power prediction based on temporal self-attention mechanism and advanced signal decomposition techniques with feature fusion

With the increasing number of distributed photovoltaic (DPV) power plants, their power prediction has become increasingly important for grid stability and energy efficiency. Challenges in DPV power prediction include the inaccessibility of high-resolution meteorological data and the difficulty in processing complex multivariate time series (MTS) with high dimensionality and long-term dependencies. To address these challenges, the universal kriging method is used to interpolate meteorological data, and a novel short-term DPV power prediction model is proposed in this paper. This model employs Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and Kernel Principal Component Analysis (KPCA) for effective signal decomposition and feature extraction from MTSs. By integrating a Gated Recurrent Unit (GRU) as a sequence encoder, together with a custom-designed temporal self-attention mechanism, it captures the intricate temporal dependencies more comprehensively than standard self-attention methods, which often rely heavily on the hidden states of Recurrent Neural Networks (RNNs). In addition, the inclusion of skip connections mitigates the risk of deep learning degradation and enhances the model's ability to fuse contextual features, resulting in a robust predictive framework that effectively adapts to temporal variations. Comparative tests on real data from DPV systems with five baseline models and ablation experiments show that our approach has higher prediction accuracy and is more robust. Compared to the best performing baseline model, the mean square error is 34.3 % lower, the mean absolute error is 5.2 % lower, and R2 reaches 0.9772, the highest among all models.