【摘要】 针对天空云层变化导致光伏功率预测精度下降，以及现有基于全天空图像方法在关键云区特征提取方面存在不足的问题，提出一种基于全天空图像关键云区特征提取的改进光伏功率预测方法。首先，提取全天空图像关键云区特征与历史光伏功率特征以捕获扰动与时序信息；随后，采用双通道长短期记忆网络（LSTM）分别编码过去与未来特征，并通过注意力机制实现高效融合；最后，引入逐步训练策略，并利用常春藤算法（IVYA）优化模型参数。实测数据表明，所提方法优于空间卷积神经网络（SCNN）、ConvNeXt、双向长短期记忆网络（Bi-LSTM）及Transformer等主流方法，其10 min预测性能取得E_MAE为0.515 kW、E_RMSE为1.489 kW、R²为0.967 8的结果。更多还原

【Abstract】 To address the degradation in photovoltaic（PV） power forecasting accuracy caused by dynamic cloud evolution and the insufficient extraction of key cloud-region features in existing all-sky image–based methods, an improved PV power prediction approach based on key cloud-region feature extraction from all-sky images is proposed.Firstly, the features of key cloud regions from all-sky images and historical PV power are extracted to capture perturbation and temporal information.Subsequently, a dual-channel long short-term memory（LSTM） network is employed to encode past and future features separately, which are then efficiently fused via an attention mechanism.Finally, a stepwise training strategy is introduced, and the Ivy Algorithm（IVYA） is used to optimize model parameters.Validation using one year of measured data demonstrates that the proposed method outperforms mainstream models including SCNN,ConvNeXt, Bi-LSTM,and Transformer, achieving a 10-minute forecasting performance with a E_MAE of 0.515 kW,a E_RMSE of 1.489 kW,and R² of 0.967 8.更多还原