【目的】针对现有预测方法难以精准刻画大型新能源基地水风光多源序列的复杂时空互补特性，导致预测精度受限的问题，提出一种基于双层互补性正则化的水风光时空联合预测方法。【方法】首先，构建季节趋势分解双通道时空Transformer（seasonal trend decomposition spatial temporal Transformer，STD-ST-Former）模型，利用双通道机制对多源原始序列进行解耦。然后，在损失函数中引入双层互补性正则项对模型进行协同优化。在系统层，基于归一化加权聚合序列构建全局互补性指标，引导模型挖掘多源互补耦合特征以强化系统整体预测能力；在拓扑层，结合余弦相似度与铰链函数表征电气相邻节点的空间相关性，并基于真实相关度的门控权重实现差异化调节，自适应修正节点预测偏差。最后，采用融合双层互补性正则项的损失函数训练优化STD-ST-Former模型，在精准捕捉时序演化特征的同时有效刻画内在互补机理，生成更精确的水风光联合预测结果。【结果】基于多源电站实验结果表明，所提方法在复杂时空相关性下展现出更高的预测精度和更稳定的预测性能，相比7种基准对比模型，相对均方根误差降幅为22.0%~56.3%、相关系数整体提升0.9%~1.4%，且准确率始终保持在96.9%以上。【结论】利用双层互补性正则化的方法能够准确还原水风光出力的真实时空演化规律，为高比例可再生能源的多源协同预测提供了可行的解决方案。

[Objective] Existing prediction methods struggle to accurately characterize the complex spatiotemporal complementarity of multi-source hydro-wind-photovoltaic series in large-scale new energy bases， thereby limiting prediction accuracy. To address this issue， this paper proposes a spatiotemporal joint prediction method based on bi-level complementarity regularization. [Methods] First， a seasonal trend decomposition spatial-temporal Transformer （STD-ST-Former） is constructed， employing a dual-channel mechanism to decouple the multi-source original sequences. Subsequently， bi-level complementarity regularization terms are introduced into the loss function to collaboratively optimize the model. At the system level， a global complementarity index is constructed based on normalized weighted aggregated sequences. This guides the model to mine multi-source complementary coupling features， thereby enhancing the overall prediction capability of the system. At the topological level， cosine similarity and a hinge function are combined to characterize the spatial correlation of electrically adjacent nodes. Furthermore， differentiated adjustments are implemented based on gating weights of real correlations to adaptively correct node prediction deviations. Finally， the STD-ST-Former model is trained and optimized using a loss function that integrates the bi-level complementarity regularization terms. This allows for the precise capturing of temporal evolution features while effectively characterizing the intrinsic complementarity mechanism， generating more accurate hydro-wind-photovoltaic joint prediction results. [Results] Experimental results based on multi-source power stations demonstrate that the proposed method exhibits higher prediction accuracy and more stable performance under complex spatiotemporal correlations. Compared to the seven baseline models， the relative Root Mean Square Error is reduced by 22.0% to 56.3%， the correlation coefficient is improved by 0.9% to 1.4% overall， and the prediction accuracy is consistently maintained above 96.9%. [Conclusions] The proposed approach based on bi-level complementarity regularization accurately reconstructs the true spatiotemporal evolutionary patterns of hydro， wind and PV power. It provides a feasible solution for multi-source collaborative forecasting in power systems with high penetration of renewable energy.