针对现有模型难以有效捕获气象因素引发的光伏发电的随机性与间歇性，导致预测精度低的问题，提出一种基于FDAM-SOFTS-TKAN网络的短期光伏功率预测方法。首先，使用皮尔逊相关系数筛选关键因素作为输入变量，采用高斯混合模型聚类算法将数据划分为晴天、多云天和阴雨天三种类型；随后，提出频域深度注意力机制（Frequency Depthwise Attention Mechanism, FDAM）对序列聚合网络进行改进优化，增强模型对长期依赖关系和变量相关性的提取能力；同时，利用TKAN（Taylor Kolmogorov-Arnold Network, TKAN）高阶非线性映射提升模型对复杂气象条件的适应性；最后与koopa和iTransformer等4种当前先进的模型进行对比实验。结果表明:所提方法在晴天、多云及阴雨天气下的MAE和RMSE分别降低18.6%~67.4%和13.7%~62.6%，充分验证了其在预测精度与复杂天气鲁棒性方面的优越性能。

To address the challenge that existing models struggle to effectively capture the randomness and intermittency of photovoltaic (PV) power generation caused by meteorological factors, leading to low prediction accuracy, this paper proposes a short-term PV power forecasting method based on the FDAM-SOFTS-TKAN network. First, the Pearson correlation coefficient is employed to screen key influencing factors as input variables, and a Gaussian mixture model (GMM) clustering algorithm is applied to classify the data into three weather types: sunny, cloudy, and rainy. Subsequently, a Frequency-Depthwise Attention Mechanism (FDAM) is proposed to enhance the sequence aggregation network, improving the model"s ability to extract long-term dependencies and variable correlations. Meanwhile, the Taylor Kolmogorov-Arnold Network (TKAN) is utilized to achieve higher-order nonlinear mapping, thereby strengthening the model"s adaptability to complex meteorological conditions. Finally, comparison experiments with four state-of-the-art models, such as koopa, iTransformer, and etc. is finished. The result demonstrate that the proposed method reduces the MAE and RMSE by 18.6%~67.4% and 13.7%~62.6%, respectively, under sunny, cloudy, and rainy conditions. These results fully validate its superior performance in prediction accuracy and robustness under complex weather scenarios.

国家自然科学基金（61863003）