A hybrid ultra-short-term photovoltaic power prediction framework integrating ant colony optimization for clustering with Bi-GRU

• The proposed hybrid framework enhances photovoltaic power prediction by scenario-based strategy. • The novel K-MACO algorithm boosts weather-scenario compactness and separation. • Kendall’s Tau derives the optimal feature subset to refine neural network inputs. • Scenario-driven Bi-GRU compresses overall error further. To address the challenges posed to stable power grid operation by the intermittency and volatility of photovoltaic (PV) power generation, this paper proposes an ultra-short-term hybrid PV power prediction model based on a logical framework integrating a scenario-based prediction strategy. This hybrid framework employs an improved Ant Colony Optimization algorithm fused with K-Means pre-clustering (K-MACO) to perform unsupervised learning on samples within the physical feature space of radiation patterns, humidity, and temperature dynamics, classifying weather scenarios into sunny, cloudy, and rainy types. Compared to derivative models, this approach improves the Dunn Index by 24 % and the Silhouette Coefficient by 34 %. Kendall’s Rank Correlation Coefficient is utilized as the core metric for feature selection to identify the optimal predictive feature subset, resulting in a 9-feature set that achieves the best predictive performance (R 2 = 0.8153). Bidirectional Gated Recurrent Unit (Bi-GRU) models are independently constructed based on these scenario labels to enable scenario-adaptive predictions. Evaluated on 15-minute resolution data from a power station, the proposed framework shows an average performance improvement of 3 %–10 % on R 2 , RMSE, and MAE metrics over baseline models such as BP, TCN, and Transformer (R 2 = 0.8187, MAE = 8.01), with the scenario-based strategy contributing an additional 4.5 % reduction in MAE (7.65 vs. 8.01). The research indicates that this synergistic design, which integrates scenario division, feature selection, and scene-adaptive prediction, significantly enhances the accuracy of PV power prediction and can provide a reliable basis for decision-making in grid dispatching with high renewable energy penetration.