【目的】针对风电集群功率预测中时空特征提取不足和多尺度特征难以有效融合的问题，提出一种时空图谱聚类（spatio-temporal graph spectral clustering，ST-SpecCluster）与时序二维化Transformer（time-series 2D-reshaping transformer， T2Dformer）的风电集群功率多尺度融合预测方法。【方法】首先，以风电场地理坐标构建空间骨架，并基于滑动时间窗口内的功率数据动态计算风电场间的皮尔逊相关系数，构建了一种能同时感知静态地理邻近性与动态运行关联性的混合时空图；通过时空图卷积网络提取深度时空特征，并融合注意力聚合机制与谱聚类划分出强相关子集群；随后，利用局部加权回归的季节与趋势分解（seasonal-trend decomposition procedure based on loess， STL）将子集群内功率及数值天气预报序列分解出趋势项、周期项和残差项。最后，设计了新的预测模型，通过快速傅里叶变换周期提取、时序二维化、Inception卷积与注意力机制的协同，实现多尺度特征联合建模。利用时序二维化Transformer分别预测各个集群的各个分量，并合并重构结果，将该方法应用于中国吉林某风电场集群。【结果】与前沿的预测方法相比，所提方法的归一化均方根误差降低了1.89%，归一化平均绝对误差降低了2.87%，决定系数提升了10.87%。【结论】所提方法为高精度风电功率预测提供了一种有效解决方案，对提升电力系统调度能力与新能源消纳具有重要实际意义。

[Objective] To address limitations in spatio-temporal feature extraction and the integration of multi-scale features in wind power cluster prediction， this paper proposes a multi-scale fusion prediction method based on spatio-temporal graph spectral clustering （ST-SpecCluster） and a time-series 2D-reshaping Ttransformer （T2Dformer）. [Methods] A spatial skeleton is constructed utilizing the geographical coordinates of wind farms. A hybrid spatio-temporal graph is built to capture both static geographical proximity and dynamic operational correlations by dynamically calculating the Pearson correlation coefficients between wind farms using power data within sliding time windows. Deep spatio-temporal features are extracted via a spatio-temporal graph convolutional network （ST-GCN）， and an attention aggregation mechanism is combined with spectral clustering to partition the wind farms into highly correlated sub-clusters. For each sub-cluster， the power and numerical weather prediction （NWP） sequences are decomposed into trend， periodic， and residual components using the seasonal-trend decomposition procedure based on loess （STL）. A novel prediction model， the T2Dformer， is then designed to jointly model multi-scale features through the collaboration of fast Fourier transform （FFT） period extraction， time-series 2D reshaping， Inception convolution， and attention mechanisms. Each component of every cluster is predicted separately， followed by aggregation and reconstruction. [Results] The proposed method was applied to a wind farm cluster in Jilin Province， China. Compared with state-of-the-art prediction methods， the proposed approach reduced the normalized root mean square error by 1.89%， reduced the normalized mean absolute error by 2.87%， and increased the coefficient of determination （R²） by 10.87%. [Conclusions] This paper provides an effective solution for high-precision wind power prediction， with practical significance for enhancing the dispatching capability of power systems and the consumption of renewable energy.