Ultra-short-term PV power prediction based on an improved hybrid model with sky image features and data two-dimensional purification

Ultra-short-term photovoltaic (PV) power prediction facilitates activities such as formulating charging and discharging strategies for energy storage systems and engaging in short-term trading in the electricity market, thereby improving system operation efficiency and economy. To enhance the accuracy and efficiency of prediction models by utilizing multi-source information, this paper proposes a novel method for 15-min-ahead PV power prediction, which is based on the improved Ensemble Empirical Mode Decomposition with extreme points as termination conditions (ET-EEMD) and Gated Recurrent Unit (GRU) integrated with sky image feature and data two-dimensional purification. The proposed method begins with a novel approach based on color quantization for extracting features from sky images, utilizing the LAB color space and K-means clustering. This approach establishes a connection between pixel color and image feature, and is lightweight, stable, and interpretable. Consequently, a data two-dimensional purification method based on Grey Relation Analysis (GRA) and Principal Component Analysis (PCA) is employed to select variables strongly correlated with PV power and eliminate coupling for better compatibility and higher computational efficiency. Then, ET-EEMD is applied to utilize the periodicity of PV power and the frequency-separation of IMFs, which is then combined with the GRU to form a hybrid prediction model, thus effectively improving the modeling efficiency while ensuring the prediction accuracy. Finally, the feasibility and superiority of the proposed method are validated through four sets of comparison experiments, demonstrating impressive performance across four evaluation metrics: NRMSE, MAPE, R2, and runtime, with values of 2.37 %, 1.23 %, 0.99, and 3215.26s, where the sky image features reduced the NRMSE by 38.3 %, data purification reduced the runtime by 26.4 %. Finally, the ET-EEMD-GRU hybrid model further reduced the NRMSE by 34.0 % compared to the single GRU, and reduced the runtime by 22.4 % compared to the classical EEMD-GRU. Therefore, the proposed prediction method achieves improvements in both prediction accuracy and training efficiency with sky image features and data purification.